2017-06-07 23:53:32 +00:00
|
|
|
//===- TailDuplicator.cpp - Duplicate blocks into predecessors' tails -----===//
|
2016-04-08 20:35:01 +00:00
|
|
|
//
|
|
|
|
// The LLVM Compiler Infrastructure
|
|
|
|
//
|
|
|
|
// This file is distributed under the University of Illinois Open Source
|
|
|
|
// License. See LICENSE.TXT for details.
|
|
|
|
//
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
//
|
|
|
|
// This utility class duplicates basic blocks ending in unconditional branches
|
|
|
|
// into the tails of their predecessors.
|
|
|
|
//
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
2017-11-08 01:01:31 +00:00
|
|
|
#include "llvm/CodeGen/TailDuplicator.h"
|
2017-06-07 23:53:32 +00:00
|
|
|
#include "llvm/ADT/DenseMap.h"
|
2016-04-08 20:35:01 +00:00
|
|
|
#include "llvm/ADT/DenseSet.h"
|
2017-11-08 01:01:31 +00:00
|
|
|
#include "llvm/ADT/STLExtras.h"
|
2016-04-08 20:35:01 +00:00
|
|
|
#include "llvm/ADT/SetVector.h"
|
2017-06-07 23:53:32 +00:00
|
|
|
#include "llvm/ADT/SmallPtrSet.h"
|
|
|
|
#include "llvm/ADT/SmallVector.h"
|
2016-04-08 20:35:01 +00:00
|
|
|
#include "llvm/ADT/Statistic.h"
|
2017-06-07 23:53:32 +00:00
|
|
|
#include "llvm/CodeGen/MachineBasicBlock.h"
|
2016-04-08 20:35:01 +00:00
|
|
|
#include "llvm/CodeGen/MachineBranchProbabilityInfo.h"
|
2017-06-07 23:53:32 +00:00
|
|
|
#include "llvm/CodeGen/MachineFunction.h"
|
|
|
|
#include "llvm/CodeGen/MachineInstr.h"
|
2016-04-08 20:35:01 +00:00
|
|
|
#include "llvm/CodeGen/MachineInstrBuilder.h"
|
2017-06-07 23:53:32 +00:00
|
|
|
#include "llvm/CodeGen/MachineOperand.h"
|
|
|
|
#include "llvm/CodeGen/MachineRegisterInfo.h"
|
|
|
|
#include "llvm/CodeGen/MachineSSAUpdater.h"
|
2017-11-08 01:01:31 +00:00
|
|
|
#include "llvm/CodeGen/TargetInstrInfo.h"
|
2017-11-17 01:07:10 +00:00
|
|
|
#include "llvm/CodeGen/TargetRegisterInfo.h"
|
|
|
|
#include "llvm/CodeGen/TargetSubtargetInfo.h"
|
2017-06-07 23:53:32 +00:00
|
|
|
#include "llvm/IR/DebugLoc.h"
|
2016-04-08 20:35:01 +00:00
|
|
|
#include "llvm/IR/Function.h"
|
|
|
|
#include "llvm/Support/CommandLine.h"
|
|
|
|
#include "llvm/Support/Debug.h"
|
|
|
|
#include "llvm/Support/ErrorHandling.h"
|
|
|
|
#include "llvm/Support/raw_ostream.h"
|
2018-01-31 15:57:57 +00:00
|
|
|
#include "llvm/Target/TargetMachine.h"
|
2017-06-07 23:53:32 +00:00
|
|
|
#include <algorithm>
|
|
|
|
#include <cassert>
|
|
|
|
#include <iterator>
|
|
|
|
#include <utility>
|
|
|
|
|
2016-04-08 20:35:01 +00:00
|
|
|
using namespace llvm;
|
|
|
|
|
|
|
|
#define DEBUG_TYPE "tailduplication"
|
|
|
|
|
|
|
|
STATISTIC(NumTails, "Number of tails duplicated");
|
|
|
|
STATISTIC(NumTailDups, "Number of tail duplicated blocks");
|
2016-06-14 19:40:10 +00:00
|
|
|
STATISTIC(NumTailDupAdded,
|
|
|
|
"Number of instructions added due to tail duplication");
|
|
|
|
STATISTIC(NumTailDupRemoved,
|
|
|
|
"Number of instructions removed due to tail duplication");
|
2016-04-08 20:35:01 +00:00
|
|
|
STATISTIC(NumDeadBlocks, "Number of dead blocks removed");
|
|
|
|
STATISTIC(NumAddedPHIs, "Number of phis added");
|
|
|
|
|
|
|
|
// Heuristic for tail duplication.
|
|
|
|
static cl::opt<unsigned> TailDuplicateSize(
|
|
|
|
"tail-dup-size",
|
|
|
|
cl::desc("Maximum instructions to consider tail duplicating"), cl::init(2),
|
|
|
|
cl::Hidden);
|
|
|
|
|
2017-06-07 23:53:32 +00:00
|
|
|
static cl::opt<unsigned> TailDupIndirectBranchSize(
|
2016-08-30 18:18:54 +00:00
|
|
|
"tail-dup-indirect-size",
|
|
|
|
cl::desc("Maximum instructions to consider tail duplicating blocks that "
|
|
|
|
"end with indirect branches."), cl::init(20),
|
|
|
|
cl::Hidden);
|
|
|
|
|
2016-04-08 20:35:01 +00:00
|
|
|
static cl::opt<bool>
|
|
|
|
TailDupVerify("tail-dup-verify",
|
|
|
|
cl::desc("Verify sanity of PHI instructions during taildup"),
|
|
|
|
cl::init(false), cl::Hidden);
|
|
|
|
|
|
|
|
static cl::opt<unsigned> TailDupLimit("tail-dup-limit", cl::init(~0U),
|
|
|
|
cl::Hidden);
|
|
|
|
|
2017-08-23 03:17:59 +00:00
|
|
|
void TailDuplicator::initMF(MachineFunction &MFin, bool PreRegAlloc,
|
2016-08-17 21:07:35 +00:00
|
|
|
const MachineBranchProbabilityInfo *MBPIin,
|
Codegen: Tail-duplicate during placement.
The tail duplication pass uses an assumed layout when making duplication
decisions. This is fine, but passes up duplication opportunities that
may arise when blocks are outlined. Because we want the updated CFG to
affect subsequent placement decisions, this change must occur during
placement.
In order to achieve this goal, TailDuplicationPass is split into a
utility class, TailDuplicator, and the pass itself. The pass delegates
nearly everything to the TailDuplicator object, except for looping over
the blocks in a function. This allows the same code to be used for tail
duplication in both places.
This change, in concert with outlining optional branches, allows
triangle shaped code to perform much better, esepecially when the
taken/untaken branches are correlated, as it creates a second spine when
the tests are small enough.
Issue from previous rollback fixed, and a new test was added for that
case as well. Issue was worklist/scheduling/taildup issue in layout.
Issue from 2nd rollback fixed, with 2 additional tests. Issue was
tail merging/loop info/tail-duplication causing issue with loops that share
a header block.
Issue with early tail-duplication of blocks that branch to a fallthrough
predecessor fixed with test case: tail-dup-branch-to-fallthrough.ll
Differential revision: https://reviews.llvm.org/D18226
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@283934 91177308-0d34-0410-b5e6-96231b3b80d8
2016-10-11 20:36:43 +00:00
|
|
|
bool LayoutModeIn, unsigned TailDupSizeIn) {
|
2016-08-25 01:37:03 +00:00
|
|
|
MF = &MFin;
|
|
|
|
TII = MF->getSubtarget().getInstrInfo();
|
|
|
|
TRI = MF->getSubtarget().getRegisterInfo();
|
|
|
|
MRI = &MF->getRegInfo();
|
2016-08-25 01:37:07 +00:00
|
|
|
MMI = &MF->getMMI();
|
2016-04-08 20:35:01 +00:00
|
|
|
MBPI = MBPIin;
|
2016-08-17 21:07:35 +00:00
|
|
|
TailDupSize = TailDupSizeIn;
|
2016-04-08 20:35:01 +00:00
|
|
|
|
|
|
|
assert(MBPI != nullptr && "Machine Branch Probability Info required");
|
|
|
|
|
Codegen: Tail-duplicate during placement.
The tail duplication pass uses an assumed layout when making duplication
decisions. This is fine, but passes up duplication opportunities that
may arise when blocks are outlined. Because we want the updated CFG to
affect subsequent placement decisions, this change must occur during
placement.
In order to achieve this goal, TailDuplicationPass is split into a
utility class, TailDuplicator, and the pass itself. The pass delegates
nearly everything to the TailDuplicator object, except for looping over
the blocks in a function. This allows the same code to be used for tail
duplication in both places.
This change, in concert with outlining optional branches, allows
triangle shaped code to perform much better, esepecially when the
taken/untaken branches are correlated, as it creates a second spine when
the tests are small enough.
Issue from previous rollback fixed, and a new test was added for that
case as well. Issue was worklist/scheduling/taildup issue in layout.
Issue from 2nd rollback fixed, with 2 additional tests. Issue was
tail merging/loop info/tail-duplication causing issue with loops that share
a header block.
Issue with early tail-duplication of blocks that branch to a fallthrough
predecessor fixed with test case: tail-dup-branch-to-fallthrough.ll
Differential revision: https://reviews.llvm.org/D18226
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@283934 91177308-0d34-0410-b5e6-96231b3b80d8
2016-10-11 20:36:43 +00:00
|
|
|
LayoutMode = LayoutModeIn;
|
2017-08-23 03:17:59 +00:00
|
|
|
this->PreRegAlloc = PreRegAlloc;
|
2016-04-08 20:35:01 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
static void VerifyPHIs(MachineFunction &MF, bool CheckExtra) {
|
|
|
|
for (MachineFunction::iterator I = ++MF.begin(), E = MF.end(); I != E; ++I) {
|
|
|
|
MachineBasicBlock *MBB = &*I;
|
|
|
|
SmallSetVector<MachineBasicBlock *, 8> Preds(MBB->pred_begin(),
|
|
|
|
MBB->pred_end());
|
|
|
|
MachineBasicBlock::iterator MI = MBB->begin();
|
|
|
|
while (MI != MBB->end()) {
|
|
|
|
if (!MI->isPHI())
|
|
|
|
break;
|
2016-08-16 20:38:05 +00:00
|
|
|
for (MachineBasicBlock *PredBB : Preds) {
|
2016-04-08 20:35:01 +00:00
|
|
|
bool Found = false;
|
|
|
|
for (unsigned i = 1, e = MI->getNumOperands(); i != e; i += 2) {
|
|
|
|
MachineBasicBlock *PHIBB = MI->getOperand(i + 1).getMBB();
|
|
|
|
if (PHIBB == PredBB) {
|
|
|
|
Found = true;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
if (!Found) {
|
2017-12-04 17:18:51 +00:00
|
|
|
dbgs() << "Malformed PHI in " << printMBBReference(*MBB) << ": "
|
|
|
|
<< *MI;
|
|
|
|
dbgs() << " missing input from predecessor "
|
|
|
|
<< printMBBReference(*PredBB) << '\n';
|
2016-04-08 20:35:01 +00:00
|
|
|
llvm_unreachable(nullptr);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
for (unsigned i = 1, e = MI->getNumOperands(); i != e; i += 2) {
|
|
|
|
MachineBasicBlock *PHIBB = MI->getOperand(i + 1).getMBB();
|
|
|
|
if (CheckExtra && !Preds.count(PHIBB)) {
|
2017-12-04 17:18:51 +00:00
|
|
|
dbgs() << "Warning: malformed PHI in " << printMBBReference(*MBB)
|
|
|
|
<< ": " << *MI;
|
|
|
|
dbgs() << " extra input from predecessor "
|
|
|
|
<< printMBBReference(*PHIBB) << '\n';
|
2016-04-08 20:35:01 +00:00
|
|
|
llvm_unreachable(nullptr);
|
|
|
|
}
|
|
|
|
if (PHIBB->getNumber() < 0) {
|
2017-12-04 17:18:51 +00:00
|
|
|
dbgs() << "Malformed PHI in " << printMBBReference(*MBB) << ": "
|
|
|
|
<< *MI;
|
|
|
|
dbgs() << " non-existing " << printMBBReference(*PHIBB) << '\n';
|
2016-04-08 20:35:01 +00:00
|
|
|
llvm_unreachable(nullptr);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
++MI;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/// Tail duplicate the block and cleanup.
|
2016-08-26 20:12:40 +00:00
|
|
|
/// \p IsSimple - return value of isSimpleBB
|
|
|
|
/// \p MBB - block to be duplicated
|
Codegen: Tail-duplicate during placement.
The tail duplication pass uses an assumed layout when making duplication
decisions. This is fine, but passes up duplication opportunities that
may arise when blocks are outlined. Because we want the updated CFG to
affect subsequent placement decisions, this change must occur during
placement.
In order to achieve this goal, TailDuplicationPass is split into a
utility class, TailDuplicator, and the pass itself. The pass delegates
nearly everything to the TailDuplicator object, except for looping over
the blocks in a function. This allows the same code to be used for tail
duplication in both places.
This change, in concert with outlining optional branches, allows
triangle shaped code to perform much better, esepecially when the
taken/untaken branches are correlated, as it creates a second spine when
the tests are small enough.
Issue from previous rollback fixed, and a new test was added for that
case as well. Issue was worklist/scheduling/taildup issue in layout.
Issue from 2nd rollback fixed, with 2 additional tests. Issue was
tail merging/loop info/tail-duplication causing issue with loops that share
a header block.
Issue with early tail-duplication of blocks that branch to a fallthrough
predecessor fixed with test case: tail-dup-branch-to-fallthrough.ll
Differential revision: https://reviews.llvm.org/D18226
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@283934 91177308-0d34-0410-b5e6-96231b3b80d8
2016-10-11 20:36:43 +00:00
|
|
|
/// \p ForcedLayoutPred - If non-null, treat this block as the layout
|
|
|
|
/// predecessor, instead of using the ordering in MF
|
2016-08-26 20:12:40 +00:00
|
|
|
/// \p DuplicatedPreds - if non-null, \p DuplicatedPreds will contain a list of
|
|
|
|
/// all Preds that received a copy of \p MBB.
|
Codegen: Tail-duplicate during placement.
The tail duplication pass uses an assumed layout when making duplication
decisions. This is fine, but passes up duplication opportunities that
may arise when blocks are outlined. Because we want the updated CFG to
affect subsequent placement decisions, this change must occur during
placement.
In order to achieve this goal, TailDuplicationPass is split into a
utility class, TailDuplicator, and the pass itself. The pass delegates
nearly everything to the TailDuplicator object, except for looping over
the blocks in a function. This allows the same code to be used for tail
duplication in both places.
This change, in concert with outlining optional branches, allows
triangle shaped code to perform much better, esepecially when the
taken/untaken branches are correlated, as it creates a second spine when
the tests are small enough.
Issue from previous rollback fixed, and a new test was added for that
case as well. Issue was worklist/scheduling/taildup issue in layout.
Issue from 2nd rollback fixed, with 2 additional tests. Issue was
tail merging/loop info/tail-duplication causing issue with loops that share
a header block.
Issue with early tail-duplication of blocks that branch to a fallthrough
predecessor fixed with test case: tail-dup-branch-to-fallthrough.ll
Differential revision: https://reviews.llvm.org/D18226
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@283934 91177308-0d34-0410-b5e6-96231b3b80d8
2016-10-11 20:36:43 +00:00
|
|
|
/// \p RemovalCallback - if non-null, called just before MBB is deleted.
|
2016-08-26 20:12:40 +00:00
|
|
|
bool TailDuplicator::tailDuplicateAndUpdate(
|
|
|
|
bool IsSimple, MachineBasicBlock *MBB,
|
Codegen: Tail-duplicate during placement.
The tail duplication pass uses an assumed layout when making duplication
decisions. This is fine, but passes up duplication opportunities that
may arise when blocks are outlined. Because we want the updated CFG to
affect subsequent placement decisions, this change must occur during
placement.
In order to achieve this goal, TailDuplicationPass is split into a
utility class, TailDuplicator, and the pass itself. The pass delegates
nearly everything to the TailDuplicator object, except for looping over
the blocks in a function. This allows the same code to be used for tail
duplication in both places.
This change, in concert with outlining optional branches, allows
triangle shaped code to perform much better, esepecially when the
taken/untaken branches are correlated, as it creates a second spine when
the tests are small enough.
Issue from previous rollback fixed, and a new test was added for that
case as well. Issue was worklist/scheduling/taildup issue in layout.
Issue from 2nd rollback fixed, with 2 additional tests. Issue was
tail merging/loop info/tail-duplication causing issue with loops that share
a header block.
Issue with early tail-duplication of blocks that branch to a fallthrough
predecessor fixed with test case: tail-dup-branch-to-fallthrough.ll
Differential revision: https://reviews.llvm.org/D18226
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@283934 91177308-0d34-0410-b5e6-96231b3b80d8
2016-10-11 20:36:43 +00:00
|
|
|
MachineBasicBlock *ForcedLayoutPred,
|
|
|
|
SmallVectorImpl<MachineBasicBlock*> *DuplicatedPreds,
|
2017-06-07 23:53:32 +00:00
|
|
|
function_ref<void(MachineBasicBlock *)> *RemovalCallback) {
|
2016-04-08 20:35:01 +00:00
|
|
|
// Save the successors list.
|
|
|
|
SmallSetVector<MachineBasicBlock *, 8> Succs(MBB->succ_begin(),
|
|
|
|
MBB->succ_end());
|
|
|
|
|
|
|
|
SmallVector<MachineBasicBlock *, 8> TDBBs;
|
|
|
|
SmallVector<MachineInstr *, 16> Copies;
|
Codegen: Tail-duplicate during placement.
The tail duplication pass uses an assumed layout when making duplication
decisions. This is fine, but passes up duplication opportunities that
may arise when blocks are outlined. Because we want the updated CFG to
affect subsequent placement decisions, this change must occur during
placement.
In order to achieve this goal, TailDuplicationPass is split into a
utility class, TailDuplicator, and the pass itself. The pass delegates
nearly everything to the TailDuplicator object, except for looping over
the blocks in a function. This allows the same code to be used for tail
duplication in both places.
This change, in concert with outlining optional branches, allows
triangle shaped code to perform much better, esepecially when the
taken/untaken branches are correlated, as it creates a second spine when
the tests are small enough.
Issue from previous rollback fixed, and a new test was added for that
case as well. Issue was worklist/scheduling/taildup issue in layout.
Issue from 2nd rollback fixed, with 2 additional tests. Issue was
tail merging/loop info/tail-duplication causing issue with loops that share
a header block.
Issue with early tail-duplication of blocks that branch to a fallthrough
predecessor fixed with test case: tail-dup-branch-to-fallthrough.ll
Differential revision: https://reviews.llvm.org/D18226
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@283934 91177308-0d34-0410-b5e6-96231b3b80d8
2016-10-11 20:36:43 +00:00
|
|
|
if (!tailDuplicate(IsSimple, MBB, ForcedLayoutPred, TDBBs, Copies))
|
2016-04-08 20:35:01 +00:00
|
|
|
return false;
|
|
|
|
|
|
|
|
++NumTails;
|
|
|
|
|
|
|
|
SmallVector<MachineInstr *, 8> NewPHIs;
|
2016-08-25 01:37:03 +00:00
|
|
|
MachineSSAUpdater SSAUpdate(*MF, &NewPHIs);
|
2016-04-08 20:35:01 +00:00
|
|
|
|
|
|
|
// TailBB's immediate successors are now successors of those predecessors
|
|
|
|
// which duplicated TailBB. Add the predecessors as sources to the PHI
|
|
|
|
// instructions.
|
|
|
|
bool isDead = MBB->pred_empty() && !MBB->hasAddressTaken();
|
|
|
|
if (PreRegAlloc)
|
|
|
|
updateSuccessorsPHIs(MBB, isDead, TDBBs, Succs);
|
|
|
|
|
|
|
|
// If it is dead, remove it.
|
|
|
|
if (isDead) {
|
2016-06-14 19:40:10 +00:00
|
|
|
NumTailDupRemoved += MBB->size();
|
Codegen: Tail-duplicate during placement.
The tail duplication pass uses an assumed layout when making duplication
decisions. This is fine, but passes up duplication opportunities that
may arise when blocks are outlined. Because we want the updated CFG to
affect subsequent placement decisions, this change must occur during
placement.
In order to achieve this goal, TailDuplicationPass is split into a
utility class, TailDuplicator, and the pass itself. The pass delegates
nearly everything to the TailDuplicator object, except for looping over
the blocks in a function. This allows the same code to be used for tail
duplication in both places.
This change, in concert with outlining optional branches, allows
triangle shaped code to perform much better, esepecially when the
taken/untaken branches are correlated, as it creates a second spine when
the tests are small enough.
Issue from previous rollback fixed, and a new test was added for that
case as well. Issue was worklist/scheduling/taildup issue in layout.
Issue from 2nd rollback fixed, with 2 additional tests. Issue was
tail merging/loop info/tail-duplication causing issue with loops that share
a header block.
Issue with early tail-duplication of blocks that branch to a fallthrough
predecessor fixed with test case: tail-dup-branch-to-fallthrough.ll
Differential revision: https://reviews.llvm.org/D18226
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@283934 91177308-0d34-0410-b5e6-96231b3b80d8
2016-10-11 20:36:43 +00:00
|
|
|
removeDeadBlock(MBB, RemovalCallback);
|
2016-04-08 20:35:01 +00:00
|
|
|
++NumDeadBlocks;
|
|
|
|
}
|
|
|
|
|
|
|
|
// Update SSA form.
|
|
|
|
if (!SSAUpdateVRs.empty()) {
|
|
|
|
for (unsigned i = 0, e = SSAUpdateVRs.size(); i != e; ++i) {
|
|
|
|
unsigned VReg = SSAUpdateVRs[i];
|
|
|
|
SSAUpdate.Initialize(VReg);
|
|
|
|
|
|
|
|
// If the original definition is still around, add it as an available
|
|
|
|
// value.
|
|
|
|
MachineInstr *DefMI = MRI->getVRegDef(VReg);
|
|
|
|
MachineBasicBlock *DefBB = nullptr;
|
|
|
|
if (DefMI) {
|
|
|
|
DefBB = DefMI->getParent();
|
|
|
|
SSAUpdate.AddAvailableValue(DefBB, VReg);
|
|
|
|
}
|
|
|
|
|
|
|
|
// Add the new vregs as available values.
|
|
|
|
DenseMap<unsigned, AvailableValsTy>::iterator LI =
|
|
|
|
SSAUpdateVals.find(VReg);
|
|
|
|
for (unsigned j = 0, ee = LI->second.size(); j != ee; ++j) {
|
|
|
|
MachineBasicBlock *SrcBB = LI->second[j].first;
|
|
|
|
unsigned SrcReg = LI->second[j].second;
|
|
|
|
SSAUpdate.AddAvailableValue(SrcBB, SrcReg);
|
|
|
|
}
|
|
|
|
|
|
|
|
// Rewrite uses that are outside of the original def's block.
|
|
|
|
MachineRegisterInfo::use_iterator UI = MRI->use_begin(VReg);
|
|
|
|
while (UI != MRI->use_end()) {
|
|
|
|
MachineOperand &UseMO = *UI;
|
|
|
|
MachineInstr *UseMI = UseMO.getParent();
|
|
|
|
++UI;
|
|
|
|
if (UseMI->isDebugValue()) {
|
|
|
|
// SSAUpdate can replace the use with an undef. That creates
|
|
|
|
// a debug instruction that is a kill.
|
|
|
|
// FIXME: Should it SSAUpdate job to delete debug instructions
|
|
|
|
// instead of replacing the use with undef?
|
|
|
|
UseMI->eraseFromParent();
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
if (UseMI->getParent() == DefBB && !UseMI->isPHI())
|
|
|
|
continue;
|
|
|
|
SSAUpdate.RewriteUse(UseMO);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
SSAUpdateVRs.clear();
|
|
|
|
SSAUpdateVals.clear();
|
|
|
|
}
|
|
|
|
|
|
|
|
// Eliminate some of the copies inserted by tail duplication to maintain
|
|
|
|
// SSA form.
|
|
|
|
for (unsigned i = 0, e = Copies.size(); i != e; ++i) {
|
|
|
|
MachineInstr *Copy = Copies[i];
|
|
|
|
if (!Copy->isCopy())
|
|
|
|
continue;
|
|
|
|
unsigned Dst = Copy->getOperand(0).getReg();
|
|
|
|
unsigned Src = Copy->getOperand(1).getReg();
|
|
|
|
if (MRI->hasOneNonDBGUse(Src) &&
|
|
|
|
MRI->constrainRegClass(Src, MRI->getRegClass(Dst))) {
|
|
|
|
// Copy is the only use. Do trivial copy propagation here.
|
|
|
|
MRI->replaceRegWith(Dst, Src);
|
|
|
|
Copy->eraseFromParent();
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
if (NewPHIs.size())
|
|
|
|
NumAddedPHIs += NewPHIs.size();
|
|
|
|
|
2016-08-26 20:12:40 +00:00
|
|
|
if (DuplicatedPreds)
|
|
|
|
*DuplicatedPreds = std::move(TDBBs);
|
|
|
|
|
2016-04-08 20:35:01 +00:00
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
|
|
|
/// Look for small blocks that are unconditionally branched to and do not fall
|
|
|
|
/// through. Tail-duplicate their instructions into their predecessors to
|
|
|
|
/// eliminate (dynamic) branches.
|
2016-08-25 01:37:03 +00:00
|
|
|
bool TailDuplicator::tailDuplicateBlocks() {
|
2016-04-08 20:35:01 +00:00
|
|
|
bool MadeChange = false;
|
|
|
|
|
|
|
|
if (PreRegAlloc && TailDupVerify) {
|
|
|
|
DEBUG(dbgs() << "\n*** Before tail-duplicating\n");
|
2016-08-25 01:37:03 +00:00
|
|
|
VerifyPHIs(*MF, true);
|
2016-04-08 20:35:01 +00:00
|
|
|
}
|
|
|
|
|
2016-08-25 01:37:03 +00:00
|
|
|
for (MachineFunction::iterator I = ++MF->begin(), E = MF->end(); I != E;) {
|
2016-04-08 20:35:01 +00:00
|
|
|
MachineBasicBlock *MBB = &*I++;
|
|
|
|
|
|
|
|
if (NumTails == TailDupLimit)
|
|
|
|
break;
|
|
|
|
|
|
|
|
bool IsSimple = isSimpleBB(MBB);
|
|
|
|
|
2016-08-25 01:37:03 +00:00
|
|
|
if (!shouldTailDuplicate(IsSimple, *MBB))
|
2016-04-08 20:35:01 +00:00
|
|
|
continue;
|
|
|
|
|
Codegen: Tail-duplicate during placement.
The tail duplication pass uses an assumed layout when making duplication
decisions. This is fine, but passes up duplication opportunities that
may arise when blocks are outlined. Because we want the updated CFG to
affect subsequent placement decisions, this change must occur during
placement.
In order to achieve this goal, TailDuplicationPass is split into a
utility class, TailDuplicator, and the pass itself. The pass delegates
nearly everything to the TailDuplicator object, except for looping over
the blocks in a function. This allows the same code to be used for tail
duplication in both places.
This change, in concert with outlining optional branches, allows
triangle shaped code to perform much better, esepecially when the
taken/untaken branches are correlated, as it creates a second spine when
the tests are small enough.
Issue from previous rollback fixed, and a new test was added for that
case as well. Issue was worklist/scheduling/taildup issue in layout.
Issue from 2nd rollback fixed, with 2 additional tests. Issue was
tail merging/loop info/tail-duplication causing issue with loops that share
a header block.
Issue with early tail-duplication of blocks that branch to a fallthrough
predecessor fixed with test case: tail-dup-branch-to-fallthrough.ll
Differential revision: https://reviews.llvm.org/D18226
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@283934 91177308-0d34-0410-b5e6-96231b3b80d8
2016-10-11 20:36:43 +00:00
|
|
|
MadeChange |= tailDuplicateAndUpdate(IsSimple, MBB, nullptr);
|
2016-04-08 20:35:01 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
if (PreRegAlloc && TailDupVerify)
|
2016-08-25 01:37:03 +00:00
|
|
|
VerifyPHIs(*MF, false);
|
2016-04-08 20:35:01 +00:00
|
|
|
|
|
|
|
return MadeChange;
|
|
|
|
}
|
|
|
|
|
|
|
|
static bool isDefLiveOut(unsigned Reg, MachineBasicBlock *BB,
|
|
|
|
const MachineRegisterInfo *MRI) {
|
|
|
|
for (MachineInstr &UseMI : MRI->use_instructions(Reg)) {
|
|
|
|
if (UseMI.isDebugValue())
|
|
|
|
continue;
|
|
|
|
if (UseMI.getParent() != BB)
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
|
|
|
static unsigned getPHISrcRegOpIdx(MachineInstr *MI, MachineBasicBlock *SrcBB) {
|
|
|
|
for (unsigned i = 1, e = MI->getNumOperands(); i != e; i += 2)
|
|
|
|
if (MI->getOperand(i + 1).getMBB() == SrcBB)
|
|
|
|
return i;
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
// Remember which registers are used by phis in this block. This is
|
|
|
|
// used to determine which registers are liveout while modifying the
|
|
|
|
// block (which is why we need to copy the information).
|
|
|
|
static void getRegsUsedByPHIs(const MachineBasicBlock &BB,
|
|
|
|
DenseSet<unsigned> *UsedByPhi) {
|
|
|
|
for (const auto &MI : BB) {
|
|
|
|
if (!MI.isPHI())
|
|
|
|
break;
|
|
|
|
for (unsigned i = 1, e = MI.getNumOperands(); i != e; i += 2) {
|
|
|
|
unsigned SrcReg = MI.getOperand(i).getReg();
|
|
|
|
UsedByPhi->insert(SrcReg);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/// Add a definition and source virtual registers pair for SSA update.
|
|
|
|
void TailDuplicator::addSSAUpdateEntry(unsigned OrigReg, unsigned NewReg,
|
|
|
|
MachineBasicBlock *BB) {
|
|
|
|
DenseMap<unsigned, AvailableValsTy>::iterator LI =
|
|
|
|
SSAUpdateVals.find(OrigReg);
|
|
|
|
if (LI != SSAUpdateVals.end())
|
|
|
|
LI->second.push_back(std::make_pair(BB, NewReg));
|
|
|
|
else {
|
|
|
|
AvailableValsTy Vals;
|
|
|
|
Vals.push_back(std::make_pair(BB, NewReg));
|
|
|
|
SSAUpdateVals.insert(std::make_pair(OrigReg, Vals));
|
|
|
|
SSAUpdateVRs.push_back(OrigReg);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/// Process PHI node in TailBB by turning it into a copy in PredBB. Remember the
|
|
|
|
/// source register that's contributed by PredBB and update SSA update map.
|
|
|
|
void TailDuplicator::processPHI(
|
|
|
|
MachineInstr *MI, MachineBasicBlock *TailBB, MachineBasicBlock *PredBB,
|
2016-04-26 18:36:34 +00:00
|
|
|
DenseMap<unsigned, RegSubRegPair> &LocalVRMap,
|
|
|
|
SmallVectorImpl<std::pair<unsigned, RegSubRegPair>> &Copies,
|
2016-04-08 20:35:01 +00:00
|
|
|
const DenseSet<unsigned> &RegsUsedByPhi, bool Remove) {
|
|
|
|
unsigned DefReg = MI->getOperand(0).getReg();
|
|
|
|
unsigned SrcOpIdx = getPHISrcRegOpIdx(MI, PredBB);
|
|
|
|
assert(SrcOpIdx && "Unable to find matching PHI source?");
|
|
|
|
unsigned SrcReg = MI->getOperand(SrcOpIdx).getReg();
|
2016-04-26 18:36:34 +00:00
|
|
|
unsigned SrcSubReg = MI->getOperand(SrcOpIdx).getSubReg();
|
2016-04-08 20:35:01 +00:00
|
|
|
const TargetRegisterClass *RC = MRI->getRegClass(DefReg);
|
2016-04-26 18:36:34 +00:00
|
|
|
LocalVRMap.insert(std::make_pair(DefReg, RegSubRegPair(SrcReg, SrcSubReg)));
|
2016-04-08 20:35:01 +00:00
|
|
|
|
|
|
|
// Insert a copy from source to the end of the block. The def register is the
|
|
|
|
// available value liveout of the block.
|
|
|
|
unsigned NewDef = MRI->createVirtualRegister(RC);
|
2016-04-26 18:36:34 +00:00
|
|
|
Copies.push_back(std::make_pair(NewDef, RegSubRegPair(SrcReg, SrcSubReg)));
|
2016-04-08 20:35:01 +00:00
|
|
|
if (isDefLiveOut(DefReg, TailBB, MRI) || RegsUsedByPhi.count(DefReg))
|
|
|
|
addSSAUpdateEntry(DefReg, NewDef, PredBB);
|
|
|
|
|
|
|
|
if (!Remove)
|
|
|
|
return;
|
|
|
|
|
|
|
|
// Remove PredBB from the PHI node.
|
|
|
|
MI->RemoveOperand(SrcOpIdx + 1);
|
|
|
|
MI->RemoveOperand(SrcOpIdx);
|
|
|
|
if (MI->getNumOperands() == 1)
|
|
|
|
MI->eraseFromParent();
|
|
|
|
}
|
|
|
|
|
|
|
|
/// Duplicate a TailBB instruction to PredBB and update
|
|
|
|
/// the source operands due to earlier PHI translation.
|
|
|
|
void TailDuplicator::duplicateInstruction(
|
|
|
|
MachineInstr *MI, MachineBasicBlock *TailBB, MachineBasicBlock *PredBB,
|
2016-04-26 18:36:34 +00:00
|
|
|
DenseMap<unsigned, RegSubRegPair> &LocalVRMap,
|
2016-04-08 20:35:01 +00:00
|
|
|
const DenseSet<unsigned> &UsedByPhi) {
|
2018-01-31 15:57:57 +00:00
|
|
|
// Allow duplication of CFI instructions.
|
|
|
|
if (MI->isCFIInstruction()) {
|
|
|
|
BuildMI(*PredBB, PredBB->end(), PredBB->findDebugLoc(PredBB->begin()),
|
|
|
|
TII->get(TargetOpcode::CFI_INSTRUCTION)).addCFIIndex(
|
|
|
|
MI->getOperand(0).getCFIIndex());
|
|
|
|
return;
|
|
|
|
}
|
2017-08-22 23:56:30 +00:00
|
|
|
MachineInstr &NewMI = TII->duplicate(*PredBB, PredBB->end(), *MI);
|
2016-04-08 20:35:01 +00:00
|
|
|
if (PreRegAlloc) {
|
2017-08-22 23:56:30 +00:00
|
|
|
for (unsigned i = 0, e = NewMI.getNumOperands(); i != e; ++i) {
|
|
|
|
MachineOperand &MO = NewMI.getOperand(i);
|
2016-04-08 20:35:01 +00:00
|
|
|
if (!MO.isReg())
|
|
|
|
continue;
|
|
|
|
unsigned Reg = MO.getReg();
|
|
|
|
if (!TargetRegisterInfo::isVirtualRegister(Reg))
|
|
|
|
continue;
|
|
|
|
if (MO.isDef()) {
|
|
|
|
const TargetRegisterClass *RC = MRI->getRegClass(Reg);
|
|
|
|
unsigned NewReg = MRI->createVirtualRegister(RC);
|
|
|
|
MO.setReg(NewReg);
|
2016-04-26 18:36:34 +00:00
|
|
|
LocalVRMap.insert(std::make_pair(Reg, RegSubRegPair(NewReg, 0)));
|
2016-04-08 20:35:01 +00:00
|
|
|
if (isDefLiveOut(Reg, TailBB, MRI) || UsedByPhi.count(Reg))
|
|
|
|
addSSAUpdateEntry(Reg, NewReg, PredBB);
|
|
|
|
} else {
|
2016-04-26 18:36:34 +00:00
|
|
|
auto VI = LocalVRMap.find(Reg);
|
2016-04-08 20:35:01 +00:00
|
|
|
if (VI != LocalVRMap.end()) {
|
2016-04-26 18:36:34 +00:00
|
|
|
// Need to make sure that the register class of the mapped register
|
|
|
|
// will satisfy the constraints of the class of the register being
|
|
|
|
// replaced.
|
|
|
|
auto *OrigRC = MRI->getRegClass(Reg);
|
|
|
|
auto *MappedRC = MRI->getRegClass(VI->second.Reg);
|
|
|
|
const TargetRegisterClass *ConstrRC;
|
|
|
|
if (VI->second.SubReg != 0) {
|
|
|
|
ConstrRC = TRI->getMatchingSuperRegClass(MappedRC, OrigRC,
|
|
|
|
VI->second.SubReg);
|
|
|
|
if (ConstrRC) {
|
|
|
|
// The actual constraining (as in "find appropriate new class")
|
|
|
|
// is done by getMatchingSuperRegClass, so now we only need to
|
|
|
|
// change the class of the mapped register.
|
|
|
|
MRI->setRegClass(VI->second.Reg, ConstrRC);
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
// For mapped registers that do not have sub-registers, simply
|
|
|
|
// restrict their class to match the original one.
|
|
|
|
ConstrRC = MRI->constrainRegClass(VI->second.Reg, OrigRC);
|
|
|
|
}
|
|
|
|
|
|
|
|
if (ConstrRC) {
|
|
|
|
// If the class constraining succeeded, we can simply replace
|
|
|
|
// the old register with the mapped one.
|
|
|
|
MO.setReg(VI->second.Reg);
|
|
|
|
// We have Reg -> VI.Reg:VI.SubReg, so if Reg is used with a
|
|
|
|
// sub-register, we need to compose the sub-register indices.
|
|
|
|
MO.setSubReg(TRI->composeSubRegIndices(MO.getSubReg(),
|
|
|
|
VI->second.SubReg));
|
|
|
|
} else {
|
|
|
|
// The direct replacement is not possible, due to failing register
|
|
|
|
// class constraints. An explicit COPY is necessary. Create one
|
|
|
|
// that can be reused
|
|
|
|
auto *NewRC = MI->getRegClassConstraint(i, TII, TRI);
|
|
|
|
if (NewRC == nullptr)
|
|
|
|
NewRC = OrigRC;
|
|
|
|
unsigned NewReg = MRI->createVirtualRegister(NewRC);
|
|
|
|
BuildMI(*PredBB, MI, MI->getDebugLoc(),
|
|
|
|
TII->get(TargetOpcode::COPY), NewReg)
|
|
|
|
.addReg(VI->second.Reg, 0, VI->second.SubReg);
|
|
|
|
LocalVRMap.erase(VI);
|
|
|
|
LocalVRMap.insert(std::make_pair(Reg, RegSubRegPair(NewReg, 0)));
|
|
|
|
MO.setReg(NewReg);
|
|
|
|
// The composed VI.Reg:VI.SubReg is replaced with NewReg, which
|
|
|
|
// is equivalent to the whole register Reg. Hence, Reg:subreg
|
|
|
|
// is same as NewReg:subreg, so keep the sub-register index
|
|
|
|
// unchanged.
|
|
|
|
}
|
2016-04-08 20:35:01 +00:00
|
|
|
// Clear any kill flags from this operand. The new register could
|
|
|
|
// have uses after this one, so kills are not valid here.
|
|
|
|
MO.setIsKill(false);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/// After FromBB is tail duplicated into its predecessor blocks, the successors
|
|
|
|
/// have gained new predecessors. Update the PHI instructions in them
|
|
|
|
/// accordingly.
|
|
|
|
void TailDuplicator::updateSuccessorsPHIs(
|
|
|
|
MachineBasicBlock *FromBB, bool isDead,
|
|
|
|
SmallVectorImpl<MachineBasicBlock *> &TDBBs,
|
|
|
|
SmallSetVector<MachineBasicBlock *, 8> &Succs) {
|
2016-08-16 20:38:05 +00:00
|
|
|
for (MachineBasicBlock *SuccBB : Succs) {
|
|
|
|
for (MachineInstr &MI : *SuccBB) {
|
|
|
|
if (!MI.isPHI())
|
2016-04-08 20:35:01 +00:00
|
|
|
break;
|
2016-08-16 20:38:05 +00:00
|
|
|
MachineInstrBuilder MIB(*FromBB->getParent(), MI);
|
2016-04-08 20:35:01 +00:00
|
|
|
unsigned Idx = 0;
|
2016-08-16 20:38:05 +00:00
|
|
|
for (unsigned i = 1, e = MI.getNumOperands(); i != e; i += 2) {
|
|
|
|
MachineOperand &MO = MI.getOperand(i + 1);
|
2016-04-08 20:35:01 +00:00
|
|
|
if (MO.getMBB() == FromBB) {
|
|
|
|
Idx = i;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
assert(Idx != 0);
|
2016-08-16 20:38:05 +00:00
|
|
|
MachineOperand &MO0 = MI.getOperand(Idx);
|
2016-04-08 20:35:01 +00:00
|
|
|
unsigned Reg = MO0.getReg();
|
|
|
|
if (isDead) {
|
|
|
|
// Folded into the previous BB.
|
|
|
|
// There could be duplicate phi source entries. FIXME: Should sdisel
|
|
|
|
// or earlier pass fixed this?
|
2016-08-16 20:38:05 +00:00
|
|
|
for (unsigned i = MI.getNumOperands() - 2; i != Idx; i -= 2) {
|
|
|
|
MachineOperand &MO = MI.getOperand(i + 1);
|
2016-04-08 20:35:01 +00:00
|
|
|
if (MO.getMBB() == FromBB) {
|
2016-08-16 20:38:05 +00:00
|
|
|
MI.RemoveOperand(i + 1);
|
|
|
|
MI.RemoveOperand(i);
|
2016-04-08 20:35:01 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
} else
|
|
|
|
Idx = 0;
|
|
|
|
|
|
|
|
// If Idx is set, the operands at Idx and Idx+1 must be removed.
|
|
|
|
// We reuse the location to avoid expensive RemoveOperand calls.
|
|
|
|
|
|
|
|
DenseMap<unsigned, AvailableValsTy>::iterator LI =
|
|
|
|
SSAUpdateVals.find(Reg);
|
|
|
|
if (LI != SSAUpdateVals.end()) {
|
|
|
|
// This register is defined in the tail block.
|
|
|
|
for (unsigned j = 0, ee = LI->second.size(); j != ee; ++j) {
|
|
|
|
MachineBasicBlock *SrcBB = LI->second[j].first;
|
|
|
|
// If we didn't duplicate a bb into a particular predecessor, we
|
|
|
|
// might still have added an entry to SSAUpdateVals to correcly
|
|
|
|
// recompute SSA. If that case, avoid adding a dummy extra argument
|
|
|
|
// this PHI.
|
|
|
|
if (!SrcBB->isSuccessor(SuccBB))
|
|
|
|
continue;
|
|
|
|
|
|
|
|
unsigned SrcReg = LI->second[j].second;
|
|
|
|
if (Idx != 0) {
|
2016-08-16 20:38:05 +00:00
|
|
|
MI.getOperand(Idx).setReg(SrcReg);
|
|
|
|
MI.getOperand(Idx + 1).setMBB(SrcBB);
|
2016-04-08 20:35:01 +00:00
|
|
|
Idx = 0;
|
|
|
|
} else {
|
|
|
|
MIB.addReg(SrcReg).addMBB(SrcBB);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
// Live in tail block, must also be live in predecessors.
|
|
|
|
for (unsigned j = 0, ee = TDBBs.size(); j != ee; ++j) {
|
|
|
|
MachineBasicBlock *SrcBB = TDBBs[j];
|
|
|
|
if (Idx != 0) {
|
2016-08-16 20:38:05 +00:00
|
|
|
MI.getOperand(Idx).setReg(Reg);
|
|
|
|
MI.getOperand(Idx + 1).setMBB(SrcBB);
|
2016-04-08 20:35:01 +00:00
|
|
|
Idx = 0;
|
|
|
|
} else {
|
|
|
|
MIB.addReg(Reg).addMBB(SrcBB);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
if (Idx != 0) {
|
2016-08-16 20:38:05 +00:00
|
|
|
MI.RemoveOperand(Idx + 1);
|
|
|
|
MI.RemoveOperand(Idx);
|
2016-04-08 20:35:01 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/// Determine if it is profitable to duplicate this block.
|
2016-08-25 01:37:03 +00:00
|
|
|
bool TailDuplicator::shouldTailDuplicate(bool IsSimple,
|
2016-04-08 20:35:01 +00:00
|
|
|
MachineBasicBlock &TailBB) {
|
Codegen: Tail-duplicate during placement.
The tail duplication pass uses an assumed layout when making duplication
decisions. This is fine, but passes up duplication opportunities that
may arise when blocks are outlined. Because we want the updated CFG to
affect subsequent placement decisions, this change must occur during
placement.
In order to achieve this goal, TailDuplicationPass is split into a
utility class, TailDuplicator, and the pass itself. The pass delegates
nearly everything to the TailDuplicator object, except for looping over
the blocks in a function. This allows the same code to be used for tail
duplication in both places.
This change, in concert with outlining optional branches, allows
triangle shaped code to perform much better, esepecially when the
taken/untaken branches are correlated, as it creates a second spine when
the tests are small enough.
Issue from previous rollback fixed, and a new test was added for that
case as well. Issue was worklist/scheduling/taildup issue in layout.
Issue from 2nd rollback fixed, with 2 additional tests. Issue was
tail merging/loop info/tail-duplication causing issue with loops that share
a header block.
Issue with early tail-duplication of blocks that branch to a fallthrough
predecessor fixed with test case: tail-dup-branch-to-fallthrough.ll
Differential revision: https://reviews.llvm.org/D18226
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@283934 91177308-0d34-0410-b5e6-96231b3b80d8
2016-10-11 20:36:43 +00:00
|
|
|
// When doing tail-duplication during layout, the block ordering is in flux,
|
|
|
|
// so canFallThrough returns a result based on incorrect information and
|
|
|
|
// should just be ignored.
|
|
|
|
if (!LayoutMode && TailBB.canFallThrough())
|
2016-04-08 20:35:01 +00:00
|
|
|
return false;
|
|
|
|
|
|
|
|
// Don't try to tail-duplicate single-block loops.
|
|
|
|
if (TailBB.isSuccessor(&TailBB))
|
|
|
|
return false;
|
|
|
|
|
|
|
|
// Set the limit on the cost to duplicate. When optimizing for size,
|
|
|
|
// duplicate only one, because one branch instruction can be eliminated to
|
|
|
|
// compensate for the duplication.
|
|
|
|
unsigned MaxDuplicateCount;
|
2016-08-17 21:07:35 +00:00
|
|
|
if (TailDupSize == 0 &&
|
|
|
|
TailDuplicateSize.getNumOccurrences() == 0 &&
|
2017-12-15 22:22:58 +00:00
|
|
|
MF->getFunction().optForSize())
|
2016-04-08 20:35:01 +00:00
|
|
|
MaxDuplicateCount = 1;
|
2016-08-17 21:07:35 +00:00
|
|
|
else if (TailDupSize == 0)
|
2016-04-08 20:35:01 +00:00
|
|
|
MaxDuplicateCount = TailDuplicateSize;
|
2016-08-17 21:07:35 +00:00
|
|
|
else
|
|
|
|
MaxDuplicateCount = TailDupSize;
|
2016-04-08 20:35:01 +00:00
|
|
|
|
2016-08-16 22:56:14 +00:00
|
|
|
// If the block to be duplicated ends in an unanalyzable fallthrough, don't
|
|
|
|
// duplicate it.
|
|
|
|
// A similar check is necessary in MachineBlockPlacement to make sure pairs of
|
|
|
|
// blocks with unanalyzable fallthrough get layed out contiguously.
|
|
|
|
MachineBasicBlock *PredTBB = nullptr, *PredFBB = nullptr;
|
|
|
|
SmallVector<MachineOperand, 4> PredCond;
|
[MachineBlockPlacement] Don't make blocks "uneditable"
Summary:
This fixes an issue with MachineBlockPlacement due to a badly timed call
to `analyzeBranch` with `AllowModify` set to true. The timeline is as
follows:
1. `MachineBlockPlacement::maybeTailDuplicateBlock` calls
`TailDup.shouldTailDuplicate` on its argument, which in turn calls
`analyzeBranch` with `AllowModify` set to true.
2. This `analyzeBranch` call edits the terminator sequence of the block
based on the physical layout of the machine function, turning an
unanalyzable non-fallthrough block to a unanalyzable fallthrough
block. Normally MBP bails out of rearranging such blocks, but this
block was unanalyzable non-fallthrough (and thus rearrangeable) the
first time MBP looked at it, and so it goes ahead and decides where
it should be placed in the function.
3. When placing this block MBP fails to analyze and thus update the
block in keeping with the new physical layout.
Concretely, before (1) we have something like:
```
LBL0:
< unknown terminator op that may branch to LBL1 >
jmp LBL1
LBL1:
... A
LBL2:
... B
```
In (2), analyze branch simplifies this to
```
LBL0:
< unknown terminator op that may branch to LBL2 >
;; jmp LBL1 <- redundant jump removed
LBL1:
... A
LBL2:
... B
```
In (3), MachineBlockPlacement goes ahead with its plan of putting LBL2
after the first block since that is profitable.
```
LBL0:
< unknown terminator op that may branch to LBL2 >
;; jmp LBL1 <- redundant jump
LBL2:
... B
LBL1:
... A
```
and the program now has incorrect behavior (we no longer fall-through
from `LBL0` to `LBL1`) because MBP can no longer edit LBL0.
There are several possible solutions, but I went with removing the teeth
off of the `analyzeBranch` calls in TailDuplicator. That makes thinking
about the result of these calls easier, and breaks nothing in the lit
test suite.
I've also added some bookkeeping to the MachineBlockPlacement pass and
used that to write an assert that would have caught this.
Reviewers: chandlerc, gberry, MatzeB, iteratee
Subscribers: mcrosier, llvm-commits
Differential Revision: https://reviews.llvm.org/D27783
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289764 91177308-0d34-0410-b5e6-96231b3b80d8
2016-12-15 05:08:57 +00:00
|
|
|
if (TII->analyzeBranch(TailBB, PredTBB, PredFBB, PredCond) &&
|
|
|
|
TailBB.canFallThrough())
|
2016-08-16 22:56:14 +00:00
|
|
|
return false;
|
|
|
|
|
2016-04-08 20:35:01 +00:00
|
|
|
// If the target has hardware branch prediction that can handle indirect
|
|
|
|
// branches, duplicating them can often make them predictable when there
|
|
|
|
// are common paths through the code. The limit needs to be high enough
|
|
|
|
// to allow undoing the effects of tail merging and other optimizations
|
|
|
|
// that rearrange the predecessors of the indirect branch.
|
|
|
|
|
|
|
|
bool HasIndirectbr = false;
|
|
|
|
if (!TailBB.empty())
|
|
|
|
HasIndirectbr = TailBB.back().isIndirectBranch();
|
|
|
|
|
|
|
|
if (HasIndirectbr && PreRegAlloc)
|
2016-08-30 18:18:54 +00:00
|
|
|
MaxDuplicateCount = TailDupIndirectBranchSize;
|
2016-04-08 20:35:01 +00:00
|
|
|
|
|
|
|
// Check the instructions in the block to determine whether tail-duplication
|
|
|
|
// is invalid or unlikely to be profitable.
|
|
|
|
unsigned InstrCount = 0;
|
|
|
|
for (MachineInstr &MI : TailBB) {
|
|
|
|
// Non-duplicable things shouldn't be tail-duplicated.
|
2018-01-31 15:57:57 +00:00
|
|
|
// CFI instructions are marked as non-duplicable, because Darwin compact
|
|
|
|
// unwind info emission can't handle multiple prologue setups. In case of
|
|
|
|
// DWARF, allow them be duplicated, so that their existence doesn't prevent
|
|
|
|
// tail duplication of some basic blocks, that would be duplicated otherwise.
|
|
|
|
if (MI.isNotDuplicable() &&
|
|
|
|
(TailBB.getParent()->getTarget().getTargetTriple().isOSDarwin() ||
|
|
|
|
!MI.isCFIInstruction()))
|
2016-04-08 20:35:01 +00:00
|
|
|
return false;
|
|
|
|
|
|
|
|
// Convergent instructions can be duplicated only if doing so doesn't add
|
|
|
|
// new control dependencies, which is what we're going to do here.
|
|
|
|
if (MI.isConvergent())
|
|
|
|
return false;
|
|
|
|
|
|
|
|
// Do not duplicate 'return' instructions if this is a pre-regalloc run.
|
|
|
|
// A return may expand into a lot more instructions (e.g. reload of callee
|
|
|
|
// saved registers) after PEI.
|
|
|
|
if (PreRegAlloc && MI.isReturn())
|
|
|
|
return false;
|
|
|
|
|
|
|
|
// Avoid duplicating calls before register allocation. Calls presents a
|
|
|
|
// barrier to register allocation so duplicating them may end up increasing
|
|
|
|
// spills.
|
|
|
|
if (PreRegAlloc && MI.isCall())
|
|
|
|
return false;
|
|
|
|
|
2018-01-31 15:57:57 +00:00
|
|
|
if (!MI.isPHI() && !MI.isMetaInstruction())
|
Revert "Correct dwarf unwind information in function epilogue for X86"
This reverts r317579, originally committed as r317100.
There is a design issue with marking CFI instructions duplicatable. Not
all targets support the CFIInstrInserter pass, and targets like Darwin
can't cope with duplicated prologue setup CFI instructions. The compact
unwind info emission fails.
When the following code is compiled for arm64 on Mac at -O3, the CFI
instructions end up getting tail duplicated, which causes compact unwind
info emission to fail:
int a, c, d, e, f, g, h, i, j, k, l, m;
void n(int o, int *b) {
if (g)
f = 0;
for (; f < o; f++) {
m = a;
if (l > j * k > i)
j = i = k = d;
h = b[c] - e;
}
}
We get assembly that looks like this:
; BB#1: ; %if.then
Lloh3:
adrp x9, _f@GOTPAGE
Lloh4:
ldr x9, [x9, _f@GOTPAGEOFF]
mov w8, wzr
Lloh5:
str wzr, [x9]
stp x20, x19, [sp, #-16]! ; 8-byte Folded Spill
.cfi_def_cfa_offset 16
.cfi_offset w19, -8
.cfi_offset w20, -16
cmp w8, w0
b.lt LBB0_3
b LBB0_7
LBB0_2: ; %entry.if.end_crit_edge
Lloh6:
adrp x8, _f@GOTPAGE
Lloh7:
ldr x8, [x8, _f@GOTPAGEOFF]
Lloh8:
ldr w8, [x8]
stp x20, x19, [sp, #-16]! ; 8-byte Folded Spill
.cfi_def_cfa_offset 16
.cfi_offset w19, -8
.cfi_offset w20, -16
cmp w8, w0
b.ge LBB0_7
LBB0_3: ; %for.body.lr.ph
Note the multiple .cfi_def* directives. Compact unwind info emission
can't handle that.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@317726 91177308-0d34-0410-b5e6-96231b3b80d8
2017-11-08 21:31:14 +00:00
|
|
|
InstrCount += 1;
|
2016-04-08 20:35:01 +00:00
|
|
|
|
|
|
|
if (InstrCount > MaxDuplicateCount)
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
|
|
|
// Check if any of the successors of TailBB has a PHI node in which the
|
|
|
|
// value corresponding to TailBB uses a subregister.
|
|
|
|
// If a phi node uses a register paired with a subregister, the actual
|
|
|
|
// "value type" of the phi may differ from the type of the register without
|
|
|
|
// any subregisters. Due to a bug, tail duplication may add a new operand
|
|
|
|
// without a necessary subregister, producing an invalid code. This is
|
|
|
|
// demonstrated by test/CodeGen/Hexagon/tail-dup-subreg-abort.ll.
|
|
|
|
// Disable tail duplication for this case for now, until the problem is
|
|
|
|
// fixed.
|
|
|
|
for (auto SB : TailBB.successors()) {
|
|
|
|
for (auto &I : *SB) {
|
|
|
|
if (!I.isPHI())
|
|
|
|
break;
|
|
|
|
unsigned Idx = getPHISrcRegOpIdx(&I, &TailBB);
|
|
|
|
assert(Idx != 0);
|
|
|
|
MachineOperand &PU = I.getOperand(Idx);
|
|
|
|
if (PU.getSubReg() != 0)
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
if (HasIndirectbr && PreRegAlloc)
|
|
|
|
return true;
|
|
|
|
|
|
|
|
if (IsSimple)
|
|
|
|
return true;
|
|
|
|
|
|
|
|
if (!PreRegAlloc)
|
|
|
|
return true;
|
|
|
|
|
|
|
|
return canCompletelyDuplicateBB(TailBB);
|
|
|
|
}
|
|
|
|
|
|
|
|
/// True if this BB has only one unconditional jump.
|
|
|
|
bool TailDuplicator::isSimpleBB(MachineBasicBlock *TailBB) {
|
|
|
|
if (TailBB->succ_size() != 1)
|
|
|
|
return false;
|
|
|
|
if (TailBB->pred_empty())
|
|
|
|
return false;
|
|
|
|
MachineBasicBlock::iterator I = TailBB->getFirstNonDebugInstr();
|
|
|
|
if (I == TailBB->end())
|
|
|
|
return true;
|
|
|
|
return I->isUnconditionalBranch();
|
|
|
|
}
|
|
|
|
|
|
|
|
static bool bothUsedInPHI(const MachineBasicBlock &A,
|
2016-06-12 15:39:02 +00:00
|
|
|
const SmallPtrSet<MachineBasicBlock *, 8> &SuccsB) {
|
2016-04-08 20:35:01 +00:00
|
|
|
for (MachineBasicBlock *BB : A.successors())
|
|
|
|
if (SuccsB.count(BB) && !BB->empty() && BB->begin()->isPHI())
|
|
|
|
return true;
|
|
|
|
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
|
|
|
bool TailDuplicator::canCompletelyDuplicateBB(MachineBasicBlock &BB) {
|
|
|
|
for (MachineBasicBlock *PredBB : BB.predecessors()) {
|
|
|
|
if (PredBB->succ_size() > 1)
|
|
|
|
return false;
|
|
|
|
|
|
|
|
MachineBasicBlock *PredTBB = nullptr, *PredFBB = nullptr;
|
|
|
|
SmallVector<MachineOperand, 4> PredCond;
|
[MachineBlockPlacement] Don't make blocks "uneditable"
Summary:
This fixes an issue with MachineBlockPlacement due to a badly timed call
to `analyzeBranch` with `AllowModify` set to true. The timeline is as
follows:
1. `MachineBlockPlacement::maybeTailDuplicateBlock` calls
`TailDup.shouldTailDuplicate` on its argument, which in turn calls
`analyzeBranch` with `AllowModify` set to true.
2. This `analyzeBranch` call edits the terminator sequence of the block
based on the physical layout of the machine function, turning an
unanalyzable non-fallthrough block to a unanalyzable fallthrough
block. Normally MBP bails out of rearranging such blocks, but this
block was unanalyzable non-fallthrough (and thus rearrangeable) the
first time MBP looked at it, and so it goes ahead and decides where
it should be placed in the function.
3. When placing this block MBP fails to analyze and thus update the
block in keeping with the new physical layout.
Concretely, before (1) we have something like:
```
LBL0:
< unknown terminator op that may branch to LBL1 >
jmp LBL1
LBL1:
... A
LBL2:
... B
```
In (2), analyze branch simplifies this to
```
LBL0:
< unknown terminator op that may branch to LBL2 >
;; jmp LBL1 <- redundant jump removed
LBL1:
... A
LBL2:
... B
```
In (3), MachineBlockPlacement goes ahead with its plan of putting LBL2
after the first block since that is profitable.
```
LBL0:
< unknown terminator op that may branch to LBL2 >
;; jmp LBL1 <- redundant jump
LBL2:
... B
LBL1:
... A
```
and the program now has incorrect behavior (we no longer fall-through
from `LBL0` to `LBL1`) because MBP can no longer edit LBL0.
There are several possible solutions, but I went with removing the teeth
off of the `analyzeBranch` calls in TailDuplicator. That makes thinking
about the result of these calls easier, and breaks nothing in the lit
test suite.
I've also added some bookkeeping to the MachineBlockPlacement pass and
used that to write an assert that would have caught this.
Reviewers: chandlerc, gberry, MatzeB, iteratee
Subscribers: mcrosier, llvm-commits
Differential Revision: https://reviews.llvm.org/D27783
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289764 91177308-0d34-0410-b5e6-96231b3b80d8
2016-12-15 05:08:57 +00:00
|
|
|
if (TII->analyzeBranch(*PredBB, PredTBB, PredFBB, PredCond))
|
2016-04-08 20:35:01 +00:00
|
|
|
return false;
|
|
|
|
|
|
|
|
if (!PredCond.empty())
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
|
|
|
bool TailDuplicator::duplicateSimpleBB(
|
|
|
|
MachineBasicBlock *TailBB, SmallVectorImpl<MachineBasicBlock *> &TDBBs,
|
|
|
|
const DenseSet<unsigned> &UsedByPhi,
|
|
|
|
SmallVectorImpl<MachineInstr *> &Copies) {
|
|
|
|
SmallPtrSet<MachineBasicBlock *, 8> Succs(TailBB->succ_begin(),
|
|
|
|
TailBB->succ_end());
|
|
|
|
SmallVector<MachineBasicBlock *, 8> Preds(TailBB->pred_begin(),
|
|
|
|
TailBB->pred_end());
|
|
|
|
bool Changed = false;
|
2016-08-16 20:38:05 +00:00
|
|
|
for (MachineBasicBlock *PredBB : Preds) {
|
2016-04-08 20:35:01 +00:00
|
|
|
if (PredBB->hasEHPadSuccessor())
|
|
|
|
continue;
|
|
|
|
|
|
|
|
if (bothUsedInPHI(*PredBB, Succs))
|
|
|
|
continue;
|
|
|
|
|
|
|
|
MachineBasicBlock *PredTBB = nullptr, *PredFBB = nullptr;
|
|
|
|
SmallVector<MachineOperand, 4> PredCond;
|
[MachineBlockPlacement] Don't make blocks "uneditable"
Summary:
This fixes an issue with MachineBlockPlacement due to a badly timed call
to `analyzeBranch` with `AllowModify` set to true. The timeline is as
follows:
1. `MachineBlockPlacement::maybeTailDuplicateBlock` calls
`TailDup.shouldTailDuplicate` on its argument, which in turn calls
`analyzeBranch` with `AllowModify` set to true.
2. This `analyzeBranch` call edits the terminator sequence of the block
based on the physical layout of the machine function, turning an
unanalyzable non-fallthrough block to a unanalyzable fallthrough
block. Normally MBP bails out of rearranging such blocks, but this
block was unanalyzable non-fallthrough (and thus rearrangeable) the
first time MBP looked at it, and so it goes ahead and decides where
it should be placed in the function.
3. When placing this block MBP fails to analyze and thus update the
block in keeping with the new physical layout.
Concretely, before (1) we have something like:
```
LBL0:
< unknown terminator op that may branch to LBL1 >
jmp LBL1
LBL1:
... A
LBL2:
... B
```
In (2), analyze branch simplifies this to
```
LBL0:
< unknown terminator op that may branch to LBL2 >
;; jmp LBL1 <- redundant jump removed
LBL1:
... A
LBL2:
... B
```
In (3), MachineBlockPlacement goes ahead with its plan of putting LBL2
after the first block since that is profitable.
```
LBL0:
< unknown terminator op that may branch to LBL2 >
;; jmp LBL1 <- redundant jump
LBL2:
... B
LBL1:
... A
```
and the program now has incorrect behavior (we no longer fall-through
from `LBL0` to `LBL1`) because MBP can no longer edit LBL0.
There are several possible solutions, but I went with removing the teeth
off of the `analyzeBranch` calls in TailDuplicator. That makes thinking
about the result of these calls easier, and breaks nothing in the lit
test suite.
I've also added some bookkeeping to the MachineBlockPlacement pass and
used that to write an assert that would have caught this.
Reviewers: chandlerc, gberry, MatzeB, iteratee
Subscribers: mcrosier, llvm-commits
Differential Revision: https://reviews.llvm.org/D27783
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289764 91177308-0d34-0410-b5e6-96231b3b80d8
2016-12-15 05:08:57 +00:00
|
|
|
if (TII->analyzeBranch(*PredBB, PredTBB, PredFBB, PredCond))
|
2016-04-08 20:35:01 +00:00
|
|
|
continue;
|
|
|
|
|
|
|
|
Changed = true;
|
|
|
|
DEBUG(dbgs() << "\nTail-duplicating into PredBB: " << *PredBB
|
|
|
|
<< "From simple Succ: " << *TailBB);
|
|
|
|
|
|
|
|
MachineBasicBlock *NewTarget = *TailBB->succ_begin();
|
2016-08-18 00:59:32 +00:00
|
|
|
MachineBasicBlock *NextBB = PredBB->getNextNode();
|
2016-04-08 20:35:01 +00:00
|
|
|
|
|
|
|
// Make PredFBB explicit.
|
|
|
|
if (PredCond.empty())
|
|
|
|
PredFBB = PredTBB;
|
|
|
|
|
|
|
|
// Make fall through explicit.
|
|
|
|
if (!PredTBB)
|
|
|
|
PredTBB = NextBB;
|
|
|
|
if (!PredFBB)
|
|
|
|
PredFBB = NextBB;
|
|
|
|
|
|
|
|
// Redirect
|
|
|
|
if (PredFBB == TailBB)
|
|
|
|
PredFBB = NewTarget;
|
|
|
|
if (PredTBB == TailBB)
|
|
|
|
PredTBB = NewTarget;
|
|
|
|
|
|
|
|
// Make the branch unconditional if possible
|
|
|
|
if (PredTBB == PredFBB) {
|
|
|
|
PredCond.clear();
|
|
|
|
PredFBB = nullptr;
|
|
|
|
}
|
|
|
|
|
|
|
|
// Avoid adding fall through branches.
|
|
|
|
if (PredFBB == NextBB)
|
|
|
|
PredFBB = nullptr;
|
|
|
|
if (PredTBB == NextBB && PredFBB == nullptr)
|
|
|
|
PredTBB = nullptr;
|
|
|
|
|
2017-02-27 19:30:01 +00:00
|
|
|
auto DL = PredBB->findBranchDebugLoc();
|
2016-09-14 20:43:16 +00:00
|
|
|
TII->removeBranch(*PredBB);
|
2016-04-08 20:35:01 +00:00
|
|
|
|
|
|
|
if (!PredBB->isSuccessor(NewTarget))
|
|
|
|
PredBB->replaceSuccessor(TailBB, NewTarget);
|
|
|
|
else {
|
|
|
|
PredBB->removeSuccessor(TailBB, true);
|
|
|
|
assert(PredBB->succ_size() <= 1);
|
|
|
|
}
|
|
|
|
|
|
|
|
if (PredTBB)
|
2017-02-27 19:30:01 +00:00
|
|
|
TII->insertBranch(*PredBB, PredTBB, PredFBB, PredCond, DL);
|
2016-04-08 20:35:01 +00:00
|
|
|
|
|
|
|
TDBBs.push_back(PredBB);
|
|
|
|
}
|
|
|
|
return Changed;
|
|
|
|
}
|
|
|
|
|
2016-07-19 23:54:21 +00:00
|
|
|
bool TailDuplicator::canTailDuplicate(MachineBasicBlock *TailBB,
|
|
|
|
MachineBasicBlock *PredBB) {
|
Codegen: Tail-duplicate during placement.
The tail duplication pass uses an assumed layout when making duplication
decisions. This is fine, but passes up duplication opportunities that
may arise when blocks are outlined. Because we want the updated CFG to
affect subsequent placement decisions, this change must occur during
placement.
In order to achieve this goal, TailDuplicationPass is split into a
utility class, TailDuplicator, and the pass itself. The pass delegates
nearly everything to the TailDuplicator object, except for looping over
the blocks in a function. This allows the same code to be used for tail
duplication in both places.
This change, in concert with outlining optional branches, allows
triangle shaped code to perform much better, esepecially when the
taken/untaken branches are correlated, as it creates a second spine when
the tests are small enough.
Issue from previous rollback fixed, and a new test was added for that
case as well. Issue was worklist/scheduling/taildup issue in layout.
Issue from 2nd rollback fixed, with 2 additional tests. Issue was
tail merging/loop info/tail-duplication causing issue with loops that share
a header block.
Issue with early tail-duplication of blocks that branch to a fallthrough
predecessor fixed with test case: tail-dup-branch-to-fallthrough.ll
Differential revision: https://reviews.llvm.org/D18226
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@283934 91177308-0d34-0410-b5e6-96231b3b80d8
2016-10-11 20:36:43 +00:00
|
|
|
// EH edges are ignored by analyzeBranch.
|
2016-07-19 23:54:21 +00:00
|
|
|
if (PredBB->succ_size() > 1)
|
|
|
|
return false;
|
|
|
|
|
2017-05-22 21:33:54 +00:00
|
|
|
MachineBasicBlock *PredTBB = nullptr, *PredFBB = nullptr;
|
2016-07-19 23:54:21 +00:00
|
|
|
SmallVector<MachineOperand, 4> PredCond;
|
[MachineBlockPlacement] Don't make blocks "uneditable"
Summary:
This fixes an issue with MachineBlockPlacement due to a badly timed call
to `analyzeBranch` with `AllowModify` set to true. The timeline is as
follows:
1. `MachineBlockPlacement::maybeTailDuplicateBlock` calls
`TailDup.shouldTailDuplicate` on its argument, which in turn calls
`analyzeBranch` with `AllowModify` set to true.
2. This `analyzeBranch` call edits the terminator sequence of the block
based on the physical layout of the machine function, turning an
unanalyzable non-fallthrough block to a unanalyzable fallthrough
block. Normally MBP bails out of rearranging such blocks, but this
block was unanalyzable non-fallthrough (and thus rearrangeable) the
first time MBP looked at it, and so it goes ahead and decides where
it should be placed in the function.
3. When placing this block MBP fails to analyze and thus update the
block in keeping with the new physical layout.
Concretely, before (1) we have something like:
```
LBL0:
< unknown terminator op that may branch to LBL1 >
jmp LBL1
LBL1:
... A
LBL2:
... B
```
In (2), analyze branch simplifies this to
```
LBL0:
< unknown terminator op that may branch to LBL2 >
;; jmp LBL1 <- redundant jump removed
LBL1:
... A
LBL2:
... B
```
In (3), MachineBlockPlacement goes ahead with its plan of putting LBL2
after the first block since that is profitable.
```
LBL0:
< unknown terminator op that may branch to LBL2 >
;; jmp LBL1 <- redundant jump
LBL2:
... B
LBL1:
... A
```
and the program now has incorrect behavior (we no longer fall-through
from `LBL0` to `LBL1`) because MBP can no longer edit LBL0.
There are several possible solutions, but I went with removing the teeth
off of the `analyzeBranch` calls in TailDuplicator. That makes thinking
about the result of these calls easier, and breaks nothing in the lit
test suite.
I've also added some bookkeeping to the MachineBlockPlacement pass and
used that to write an assert that would have caught this.
Reviewers: chandlerc, gberry, MatzeB, iteratee
Subscribers: mcrosier, llvm-commits
Differential Revision: https://reviews.llvm.org/D27783
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289764 91177308-0d34-0410-b5e6-96231b3b80d8
2016-12-15 05:08:57 +00:00
|
|
|
if (TII->analyzeBranch(*PredBB, PredTBB, PredFBB, PredCond))
|
2016-07-19 23:54:21 +00:00
|
|
|
return false;
|
|
|
|
if (!PredCond.empty())
|
|
|
|
return false;
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
2016-04-08 20:35:01 +00:00
|
|
|
/// If it is profitable, duplicate TailBB's contents in each
|
|
|
|
/// of its predecessors.
|
Codegen: Tail-duplicate during placement.
The tail duplication pass uses an assumed layout when making duplication
decisions. This is fine, but passes up duplication opportunities that
may arise when blocks are outlined. Because we want the updated CFG to
affect subsequent placement decisions, this change must occur during
placement.
In order to achieve this goal, TailDuplicationPass is split into a
utility class, TailDuplicator, and the pass itself. The pass delegates
nearly everything to the TailDuplicator object, except for looping over
the blocks in a function. This allows the same code to be used for tail
duplication in both places.
This change, in concert with outlining optional branches, allows
triangle shaped code to perform much better, esepecially when the
taken/untaken branches are correlated, as it creates a second spine when
the tests are small enough.
Issue from previous rollback fixed, and a new test was added for that
case as well. Issue was worklist/scheduling/taildup issue in layout.
Issue from 2nd rollback fixed, with 2 additional tests. Issue was
tail merging/loop info/tail-duplication causing issue with loops that share
a header block.
Issue with early tail-duplication of blocks that branch to a fallthrough
predecessor fixed with test case: tail-dup-branch-to-fallthrough.ll
Differential revision: https://reviews.llvm.org/D18226
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@283934 91177308-0d34-0410-b5e6-96231b3b80d8
2016-10-11 20:36:43 +00:00
|
|
|
/// \p IsSimple result of isSimpleBB
|
|
|
|
/// \p TailBB Block to be duplicated.
|
|
|
|
/// \p ForcedLayoutPred When non-null, use this block as the layout predecessor
|
|
|
|
/// instead of the previous block in MF's order.
|
|
|
|
/// \p TDBBs A vector to keep track of all blocks tail-duplicated
|
|
|
|
/// into.
|
|
|
|
/// \p Copies A vector of copy instructions inserted. Used later to
|
|
|
|
/// walk all the inserted copies and remove redundant ones.
|
2016-08-25 01:37:03 +00:00
|
|
|
bool TailDuplicator::tailDuplicate(bool IsSimple, MachineBasicBlock *TailBB,
|
Codegen: Tail-duplicate during placement.
The tail duplication pass uses an assumed layout when making duplication
decisions. This is fine, but passes up duplication opportunities that
may arise when blocks are outlined. Because we want the updated CFG to
affect subsequent placement decisions, this change must occur during
placement.
In order to achieve this goal, TailDuplicationPass is split into a
utility class, TailDuplicator, and the pass itself. The pass delegates
nearly everything to the TailDuplicator object, except for looping over
the blocks in a function. This allows the same code to be used for tail
duplication in both places.
This change, in concert with outlining optional branches, allows
triangle shaped code to perform much better, esepecially when the
taken/untaken branches are correlated, as it creates a second spine when
the tests are small enough.
Issue from previous rollback fixed, and a new test was added for that
case as well. Issue was worklist/scheduling/taildup issue in layout.
Issue from 2nd rollback fixed, with 2 additional tests. Issue was
tail merging/loop info/tail-duplication causing issue with loops that share
a header block.
Issue with early tail-duplication of blocks that branch to a fallthrough
predecessor fixed with test case: tail-dup-branch-to-fallthrough.ll
Differential revision: https://reviews.llvm.org/D18226
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@283934 91177308-0d34-0410-b5e6-96231b3b80d8
2016-10-11 20:36:43 +00:00
|
|
|
MachineBasicBlock *ForcedLayoutPred,
|
2016-04-08 20:35:01 +00:00
|
|
|
SmallVectorImpl<MachineBasicBlock *> &TDBBs,
|
|
|
|
SmallVectorImpl<MachineInstr *> &Copies) {
|
2017-12-04 17:18:51 +00:00
|
|
|
DEBUG(dbgs() << "\n*** Tail-duplicating " << printMBBReference(*TailBB)
|
|
|
|
<< '\n');
|
2016-04-08 20:35:01 +00:00
|
|
|
|
|
|
|
DenseSet<unsigned> UsedByPhi;
|
|
|
|
getRegsUsedByPHIs(*TailBB, &UsedByPhi);
|
|
|
|
|
|
|
|
if (IsSimple)
|
|
|
|
return duplicateSimpleBB(TailBB, TDBBs, UsedByPhi, Copies);
|
|
|
|
|
|
|
|
// Iterate through all the unique predecessors and tail-duplicate this
|
|
|
|
// block into them, if possible. Copying the list ahead of time also
|
|
|
|
// avoids trouble with the predecessor list reallocating.
|
|
|
|
bool Changed = false;
|
|
|
|
SmallSetVector<MachineBasicBlock *, 8> Preds(TailBB->pred_begin(),
|
|
|
|
TailBB->pred_end());
|
2016-08-16 20:38:05 +00:00
|
|
|
for (MachineBasicBlock *PredBB : Preds) {
|
2016-04-08 20:35:01 +00:00
|
|
|
assert(TailBB != PredBB &&
|
|
|
|
"Single-block loop should have been rejected earlier!");
|
|
|
|
|
2016-07-19 23:54:21 +00:00
|
|
|
if (!canTailDuplicate(TailBB, PredBB))
|
2016-04-08 20:35:01 +00:00
|
|
|
continue;
|
2016-07-19 23:54:21 +00:00
|
|
|
|
2016-04-08 20:35:01 +00:00
|
|
|
// Don't duplicate into a fall-through predecessor (at least for now).
|
Codegen: Tail-duplicate during placement.
The tail duplication pass uses an assumed layout when making duplication
decisions. This is fine, but passes up duplication opportunities that
may arise when blocks are outlined. Because we want the updated CFG to
affect subsequent placement decisions, this change must occur during
placement.
In order to achieve this goal, TailDuplicationPass is split into a
utility class, TailDuplicator, and the pass itself. The pass delegates
nearly everything to the TailDuplicator object, except for looping over
the blocks in a function. This allows the same code to be used for tail
duplication in both places.
This change, in concert with outlining optional branches, allows
triangle shaped code to perform much better, esepecially when the
taken/untaken branches are correlated, as it creates a second spine when
the tests are small enough.
Issue from previous rollback fixed, and a new test was added for that
case as well. Issue was worklist/scheduling/taildup issue in layout.
Issue from 2nd rollback fixed, with 2 additional tests. Issue was
tail merging/loop info/tail-duplication causing issue with loops that share
a header block.
Issue with early tail-duplication of blocks that branch to a fallthrough
predecessor fixed with test case: tail-dup-branch-to-fallthrough.ll
Differential revision: https://reviews.llvm.org/D18226
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@283934 91177308-0d34-0410-b5e6-96231b3b80d8
2016-10-11 20:36:43 +00:00
|
|
|
bool IsLayoutSuccessor = false;
|
|
|
|
if (ForcedLayoutPred)
|
|
|
|
IsLayoutSuccessor = (ForcedLayoutPred == PredBB);
|
|
|
|
else if (PredBB->isLayoutSuccessor(TailBB) && PredBB->canFallThrough())
|
|
|
|
IsLayoutSuccessor = true;
|
|
|
|
if (IsLayoutSuccessor)
|
2016-04-08 20:35:01 +00:00
|
|
|
continue;
|
|
|
|
|
|
|
|
DEBUG(dbgs() << "\nTail-duplicating into PredBB: " << *PredBB
|
|
|
|
<< "From Succ: " << *TailBB);
|
|
|
|
|
|
|
|
TDBBs.push_back(PredBB);
|
|
|
|
|
|
|
|
// Remove PredBB's unconditional branch.
|
2016-09-14 20:43:16 +00:00
|
|
|
TII->removeBranch(*PredBB);
|
2016-04-08 20:35:01 +00:00
|
|
|
|
|
|
|
// Clone the contents of TailBB into PredBB.
|
2016-04-26 18:36:34 +00:00
|
|
|
DenseMap<unsigned, RegSubRegPair> LocalVRMap;
|
|
|
|
SmallVector<std::pair<unsigned, RegSubRegPair>, 4> CopyInfos;
|
2017-08-22 23:56:30 +00:00
|
|
|
for (MachineBasicBlock::iterator I = TailBB->begin(), E = TailBB->end();
|
|
|
|
I != E; /* empty */) {
|
2016-04-08 20:35:01 +00:00
|
|
|
MachineInstr *MI = &*I;
|
|
|
|
++I;
|
|
|
|
if (MI->isPHI()) {
|
|
|
|
// Replace the uses of the def of the PHI with the register coming
|
|
|
|
// from PredBB.
|
|
|
|
processPHI(MI, TailBB, PredBB, LocalVRMap, CopyInfos, UsedByPhi, true);
|
|
|
|
} else {
|
|
|
|
// Replace def of virtual registers with new registers, and update
|
|
|
|
// uses with PHI source register or the new registers.
|
2016-08-25 01:37:03 +00:00
|
|
|
duplicateInstruction(MI, TailBB, PredBB, LocalVRMap, UsedByPhi);
|
2016-04-08 20:35:01 +00:00
|
|
|
}
|
|
|
|
}
|
2016-04-26 18:36:34 +00:00
|
|
|
appendCopies(PredBB, CopyInfos, Copies);
|
2016-04-08 20:35:01 +00:00
|
|
|
|
|
|
|
// Simplify
|
2017-05-22 21:33:54 +00:00
|
|
|
MachineBasicBlock *PredTBB = nullptr, *PredFBB = nullptr;
|
2016-07-19 23:54:21 +00:00
|
|
|
SmallVector<MachineOperand, 4> PredCond;
|
[MachineBlockPlacement] Don't make blocks "uneditable"
Summary:
This fixes an issue with MachineBlockPlacement due to a badly timed call
to `analyzeBranch` with `AllowModify` set to true. The timeline is as
follows:
1. `MachineBlockPlacement::maybeTailDuplicateBlock` calls
`TailDup.shouldTailDuplicate` on its argument, which in turn calls
`analyzeBranch` with `AllowModify` set to true.
2. This `analyzeBranch` call edits the terminator sequence of the block
based on the physical layout of the machine function, turning an
unanalyzable non-fallthrough block to a unanalyzable fallthrough
block. Normally MBP bails out of rearranging such blocks, but this
block was unanalyzable non-fallthrough (and thus rearrangeable) the
first time MBP looked at it, and so it goes ahead and decides where
it should be placed in the function.
3. When placing this block MBP fails to analyze and thus update the
block in keeping with the new physical layout.
Concretely, before (1) we have something like:
```
LBL0:
< unknown terminator op that may branch to LBL1 >
jmp LBL1
LBL1:
... A
LBL2:
... B
```
In (2), analyze branch simplifies this to
```
LBL0:
< unknown terminator op that may branch to LBL2 >
;; jmp LBL1 <- redundant jump removed
LBL1:
... A
LBL2:
... B
```
In (3), MachineBlockPlacement goes ahead with its plan of putting LBL2
after the first block since that is profitable.
```
LBL0:
< unknown terminator op that may branch to LBL2 >
;; jmp LBL1 <- redundant jump
LBL2:
... B
LBL1:
... A
```
and the program now has incorrect behavior (we no longer fall-through
from `LBL0` to `LBL1`) because MBP can no longer edit LBL0.
There are several possible solutions, but I went with removing the teeth
off of the `analyzeBranch` calls in TailDuplicator. That makes thinking
about the result of these calls easier, and breaks nothing in the lit
test suite.
I've also added some bookkeeping to the MachineBlockPlacement pass and
used that to write an assert that would have caught this.
Reviewers: chandlerc, gberry, MatzeB, iteratee
Subscribers: mcrosier, llvm-commits
Differential Revision: https://reviews.llvm.org/D27783
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289764 91177308-0d34-0410-b5e6-96231b3b80d8
2016-12-15 05:08:57 +00:00
|
|
|
TII->analyzeBranch(*PredBB, PredTBB, PredFBB, PredCond);
|
2016-04-08 20:35:01 +00:00
|
|
|
|
2016-06-14 19:40:10 +00:00
|
|
|
NumTailDupAdded += TailBB->size() - 1; // subtract one for removed branch
|
2016-04-08 20:35:01 +00:00
|
|
|
|
|
|
|
// Update the CFG.
|
|
|
|
PredBB->removeSuccessor(PredBB->succ_begin());
|
|
|
|
assert(PredBB->succ_empty() &&
|
|
|
|
"TailDuplicate called on block with multiple successors!");
|
2016-08-16 20:38:05 +00:00
|
|
|
for (MachineBasicBlock *Succ : TailBB->successors())
|
|
|
|
PredBB->addSuccessor(Succ, MBPI->getEdgeProbability(TailBB, Succ));
|
2016-04-08 20:35:01 +00:00
|
|
|
|
|
|
|
Changed = true;
|
|
|
|
++NumTailDups;
|
|
|
|
}
|
|
|
|
|
|
|
|
// If TailBB was duplicated into all its predecessors except for the prior
|
|
|
|
// block, which falls through unconditionally, move the contents of this
|
|
|
|
// block into the prior block.
|
Codegen: Tail-duplicate during placement.
The tail duplication pass uses an assumed layout when making duplication
decisions. This is fine, but passes up duplication opportunities that
may arise when blocks are outlined. Because we want the updated CFG to
affect subsequent placement decisions, this change must occur during
placement.
In order to achieve this goal, TailDuplicationPass is split into a
utility class, TailDuplicator, and the pass itself. The pass delegates
nearly everything to the TailDuplicator object, except for looping over
the blocks in a function. This allows the same code to be used for tail
duplication in both places.
This change, in concert with outlining optional branches, allows
triangle shaped code to perform much better, esepecially when the
taken/untaken branches are correlated, as it creates a second spine when
the tests are small enough.
Issue from previous rollback fixed, and a new test was added for that
case as well. Issue was worklist/scheduling/taildup issue in layout.
Issue from 2nd rollback fixed, with 2 additional tests. Issue was
tail merging/loop info/tail-duplication causing issue with loops that share
a header block.
Issue with early tail-duplication of blocks that branch to a fallthrough
predecessor fixed with test case: tail-dup-branch-to-fallthrough.ll
Differential revision: https://reviews.llvm.org/D18226
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@283934 91177308-0d34-0410-b5e6-96231b3b80d8
2016-10-11 20:36:43 +00:00
|
|
|
MachineBasicBlock *PrevBB = ForcedLayoutPred;
|
|
|
|
if (!PrevBB)
|
|
|
|
PrevBB = &*std::prev(TailBB->getIterator());
|
2016-04-08 20:35:01 +00:00
|
|
|
MachineBasicBlock *PriorTBB = nullptr, *PriorFBB = nullptr;
|
|
|
|
SmallVector<MachineOperand, 4> PriorCond;
|
|
|
|
// This has to check PrevBB->succ_size() because EH edges are ignored by
|
Codegen: Tail-duplicate during placement.
The tail duplication pass uses an assumed layout when making duplication
decisions. This is fine, but passes up duplication opportunities that
may arise when blocks are outlined. Because we want the updated CFG to
affect subsequent placement decisions, this change must occur during
placement.
In order to achieve this goal, TailDuplicationPass is split into a
utility class, TailDuplicator, and the pass itself. The pass delegates
nearly everything to the TailDuplicator object, except for looping over
the blocks in a function. This allows the same code to be used for tail
duplication in both places.
This change, in concert with outlining optional branches, allows
triangle shaped code to perform much better, esepecially when the
taken/untaken branches are correlated, as it creates a second spine when
the tests are small enough.
Issue from previous rollback fixed, and a new test was added for that
case as well. Issue was worklist/scheduling/taildup issue in layout.
Issue from 2nd rollback fixed, with 2 additional tests. Issue was
tail merging/loop info/tail-duplication causing issue with loops that share
a header block.
Issue with early tail-duplication of blocks that branch to a fallthrough
predecessor fixed with test case: tail-dup-branch-to-fallthrough.ll
Differential revision: https://reviews.llvm.org/D18226
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@283934 91177308-0d34-0410-b5e6-96231b3b80d8
2016-10-11 20:36:43 +00:00
|
|
|
// analyzeBranch.
|
2016-04-08 20:35:01 +00:00
|
|
|
if (PrevBB->succ_size() == 1 &&
|
2016-07-20 00:01:51 +00:00
|
|
|
// Layout preds are not always CFG preds. Check.
|
|
|
|
*PrevBB->succ_begin() == TailBB &&
|
[MachineBlockPlacement] Don't make blocks "uneditable"
Summary:
This fixes an issue with MachineBlockPlacement due to a badly timed call
to `analyzeBranch` with `AllowModify` set to true. The timeline is as
follows:
1. `MachineBlockPlacement::maybeTailDuplicateBlock` calls
`TailDup.shouldTailDuplicate` on its argument, which in turn calls
`analyzeBranch` with `AllowModify` set to true.
2. This `analyzeBranch` call edits the terminator sequence of the block
based on the physical layout of the machine function, turning an
unanalyzable non-fallthrough block to a unanalyzable fallthrough
block. Normally MBP bails out of rearranging such blocks, but this
block was unanalyzable non-fallthrough (and thus rearrangeable) the
first time MBP looked at it, and so it goes ahead and decides where
it should be placed in the function.
3. When placing this block MBP fails to analyze and thus update the
block in keeping with the new physical layout.
Concretely, before (1) we have something like:
```
LBL0:
< unknown terminator op that may branch to LBL1 >
jmp LBL1
LBL1:
... A
LBL2:
... B
```
In (2), analyze branch simplifies this to
```
LBL0:
< unknown terminator op that may branch to LBL2 >
;; jmp LBL1 <- redundant jump removed
LBL1:
... A
LBL2:
... B
```
In (3), MachineBlockPlacement goes ahead with its plan of putting LBL2
after the first block since that is profitable.
```
LBL0:
< unknown terminator op that may branch to LBL2 >
;; jmp LBL1 <- redundant jump
LBL2:
... B
LBL1:
... A
```
and the program now has incorrect behavior (we no longer fall-through
from `LBL0` to `LBL1`) because MBP can no longer edit LBL0.
There are several possible solutions, but I went with removing the teeth
off of the `analyzeBranch` calls in TailDuplicator. That makes thinking
about the result of these calls easier, and breaks nothing in the lit
test suite.
I've also added some bookkeeping to the MachineBlockPlacement pass and
used that to write an assert that would have caught this.
Reviewers: chandlerc, gberry, MatzeB, iteratee
Subscribers: mcrosier, llvm-commits
Differential Revision: https://reviews.llvm.org/D27783
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@289764 91177308-0d34-0410-b5e6-96231b3b80d8
2016-12-15 05:08:57 +00:00
|
|
|
!TII->analyzeBranch(*PrevBB, PriorTBB, PriorFBB, PriorCond) &&
|
Codegen: Tail-duplicate during placement.
The tail duplication pass uses an assumed layout when making duplication
decisions. This is fine, but passes up duplication opportunities that
may arise when blocks are outlined. Because we want the updated CFG to
affect subsequent placement decisions, this change must occur during
placement.
In order to achieve this goal, TailDuplicationPass is split into a
utility class, TailDuplicator, and the pass itself. The pass delegates
nearly everything to the TailDuplicator object, except for looping over
the blocks in a function. This allows the same code to be used for tail
duplication in both places.
This change, in concert with outlining optional branches, allows
triangle shaped code to perform much better, esepecially when the
taken/untaken branches are correlated, as it creates a second spine when
the tests are small enough.
Issue from previous rollback fixed, and a new test was added for that
case as well. Issue was worklist/scheduling/taildup issue in layout.
Issue from 2nd rollback fixed, with 2 additional tests. Issue was
tail merging/loop info/tail-duplication causing issue with loops that share
a header block.
Issue with early tail-duplication of blocks that branch to a fallthrough
predecessor fixed with test case: tail-dup-branch-to-fallthrough.ll
Differential revision: https://reviews.llvm.org/D18226
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@283934 91177308-0d34-0410-b5e6-96231b3b80d8
2016-10-11 20:36:43 +00:00
|
|
|
PriorCond.empty() &&
|
|
|
|
(!PriorTBB || PriorTBB == TailBB) &&
|
|
|
|
TailBB->pred_size() == 1 &&
|
2016-04-08 20:35:01 +00:00
|
|
|
!TailBB->hasAddressTaken()) {
|
|
|
|
DEBUG(dbgs() << "\nMerging into block: " << *PrevBB
|
|
|
|
<< "From MBB: " << *TailBB);
|
Codegen: Tail-duplicate during placement.
The tail duplication pass uses an assumed layout when making duplication
decisions. This is fine, but passes up duplication opportunities that
may arise when blocks are outlined. Because we want the updated CFG to
affect subsequent placement decisions, this change must occur during
placement.
In order to achieve this goal, TailDuplicationPass is split into a
utility class, TailDuplicator, and the pass itself. The pass delegates
nearly everything to the TailDuplicator object, except for looping over
the blocks in a function. This allows the same code to be used for tail
duplication in both places.
This change, in concert with outlining optional branches, allows
triangle shaped code to perform much better, esepecially when the
taken/untaken branches are correlated, as it creates a second spine when
the tests are small enough.
Issue from previous rollback fixed, and a new test was added for that
case as well. Issue was worklist/scheduling/taildup issue in layout.
Issue from 2nd rollback fixed, with 2 additional tests. Issue was
tail merging/loop info/tail-duplication causing issue with loops that share
a header block.
Issue with early tail-duplication of blocks that branch to a fallthrough
predecessor fixed with test case: tail-dup-branch-to-fallthrough.ll
Differential revision: https://reviews.llvm.org/D18226
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@283934 91177308-0d34-0410-b5e6-96231b3b80d8
2016-10-11 20:36:43 +00:00
|
|
|
// There may be a branch to the layout successor. This is unlikely but it
|
|
|
|
// happens. The correct thing to do is to remove the branch before
|
|
|
|
// duplicating the instructions in all cases.
|
|
|
|
TII->removeBranch(*PrevBB);
|
2016-04-08 20:35:01 +00:00
|
|
|
if (PreRegAlloc) {
|
2016-04-26 18:36:34 +00:00
|
|
|
DenseMap<unsigned, RegSubRegPair> LocalVRMap;
|
|
|
|
SmallVector<std::pair<unsigned, RegSubRegPair>, 4> CopyInfos;
|
2016-04-08 20:35:01 +00:00
|
|
|
MachineBasicBlock::iterator I = TailBB->begin();
|
|
|
|
// Process PHI instructions first.
|
|
|
|
while (I != TailBB->end() && I->isPHI()) {
|
|
|
|
// Replace the uses of the def of the PHI with the register coming
|
|
|
|
// from PredBB.
|
|
|
|
MachineInstr *MI = &*I++;
|
|
|
|
processPHI(MI, TailBB, PrevBB, LocalVRMap, CopyInfos, UsedByPhi, true);
|
|
|
|
}
|
|
|
|
|
|
|
|
// Now copy the non-PHI instructions.
|
|
|
|
while (I != TailBB->end()) {
|
|
|
|
// Replace def of virtual registers with new registers, and update
|
|
|
|
// uses with PHI source register or the new registers.
|
|
|
|
MachineInstr *MI = &*I++;
|
|
|
|
assert(!MI->isBundle() && "Not expecting bundles before regalloc!");
|
2016-08-25 01:37:03 +00:00
|
|
|
duplicateInstruction(MI, TailBB, PrevBB, LocalVRMap, UsedByPhi);
|
2016-04-08 20:35:01 +00:00
|
|
|
MI->eraseFromParent();
|
|
|
|
}
|
2016-04-26 18:36:34 +00:00
|
|
|
appendCopies(PrevBB, CopyInfos, Copies);
|
2016-04-08 20:35:01 +00:00
|
|
|
} else {
|
Codegen: Tail-duplicate during placement.
The tail duplication pass uses an assumed layout when making duplication
decisions. This is fine, but passes up duplication opportunities that
may arise when blocks are outlined. Because we want the updated CFG to
affect subsequent placement decisions, this change must occur during
placement.
In order to achieve this goal, TailDuplicationPass is split into a
utility class, TailDuplicator, and the pass itself. The pass delegates
nearly everything to the TailDuplicator object, except for looping over
the blocks in a function. This allows the same code to be used for tail
duplication in both places.
This change, in concert with outlining optional branches, allows
triangle shaped code to perform much better, esepecially when the
taken/untaken branches are correlated, as it creates a second spine when
the tests are small enough.
Issue from previous rollback fixed, and a new test was added for that
case as well. Issue was worklist/scheduling/taildup issue in layout.
Issue from 2nd rollback fixed, with 2 additional tests. Issue was
tail merging/loop info/tail-duplication causing issue with loops that share
a header block.
Issue with early tail-duplication of blocks that branch to a fallthrough
predecessor fixed with test case: tail-dup-branch-to-fallthrough.ll
Differential revision: https://reviews.llvm.org/D18226
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@283934 91177308-0d34-0410-b5e6-96231b3b80d8
2016-10-11 20:36:43 +00:00
|
|
|
TII->removeBranch(*PrevBB);
|
2016-04-08 20:35:01 +00:00
|
|
|
// No PHIs to worry about, just splice the instructions over.
|
|
|
|
PrevBB->splice(PrevBB->end(), TailBB, TailBB->begin(), TailBB->end());
|
|
|
|
}
|
|
|
|
PrevBB->removeSuccessor(PrevBB->succ_begin());
|
|
|
|
assert(PrevBB->succ_empty());
|
|
|
|
PrevBB->transferSuccessors(TailBB);
|
|
|
|
TDBBs.push_back(PrevBB);
|
|
|
|
Changed = true;
|
|
|
|
}
|
|
|
|
|
|
|
|
// If this is after register allocation, there are no phis to fix.
|
|
|
|
if (!PreRegAlloc)
|
|
|
|
return Changed;
|
|
|
|
|
|
|
|
// If we made no changes so far, we are safe.
|
|
|
|
if (!Changed)
|
|
|
|
return Changed;
|
|
|
|
|
|
|
|
// Handle the nasty case in that we duplicated a block that is part of a loop
|
|
|
|
// into some but not all of its predecessors. For example:
|
|
|
|
// 1 -> 2 <-> 3 |
|
|
|
|
// \ |
|
|
|
|
// \---> rest |
|
|
|
|
// if we duplicate 2 into 1 but not into 3, we end up with
|
|
|
|
// 12 -> 3 <-> 2 -> rest |
|
|
|
|
// \ / |
|
|
|
|
// \----->-----/ |
|
|
|
|
// If there was a "var = phi(1, 3)" in 2, it has to be ultimately replaced
|
|
|
|
// with a phi in 3 (which now dominates 2).
|
|
|
|
// What we do here is introduce a copy in 3 of the register defined by the
|
|
|
|
// phi, just like when we are duplicating 2 into 3, but we don't copy any
|
|
|
|
// real instructions or remove the 3 -> 2 edge from the phi in 2.
|
2016-08-16 20:38:05 +00:00
|
|
|
for (MachineBasicBlock *PredBB : Preds) {
|
2016-08-11 22:21:41 +00:00
|
|
|
if (is_contained(TDBBs, PredBB))
|
2016-04-08 20:35:01 +00:00
|
|
|
continue;
|
|
|
|
|
|
|
|
// EH edges
|
|
|
|
if (PredBB->succ_size() != 1)
|
|
|
|
continue;
|
|
|
|
|
2016-04-26 18:36:34 +00:00
|
|
|
DenseMap<unsigned, RegSubRegPair> LocalVRMap;
|
|
|
|
SmallVector<std::pair<unsigned, RegSubRegPair>, 4> CopyInfos;
|
2016-04-08 20:35:01 +00:00
|
|
|
MachineBasicBlock::iterator I = TailBB->begin();
|
|
|
|
// Process PHI instructions first.
|
|
|
|
while (I != TailBB->end() && I->isPHI()) {
|
|
|
|
// Replace the uses of the def of the PHI with the register coming
|
|
|
|
// from PredBB.
|
|
|
|
MachineInstr *MI = &*I++;
|
|
|
|
processPHI(MI, TailBB, PredBB, LocalVRMap, CopyInfos, UsedByPhi, false);
|
|
|
|
}
|
2016-04-26 18:36:34 +00:00
|
|
|
appendCopies(PredBB, CopyInfos, Copies);
|
2016-04-08 20:35:01 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
return Changed;
|
|
|
|
}
|
|
|
|
|
2016-04-26 18:36:34 +00:00
|
|
|
/// At the end of the block \p MBB generate COPY instructions between registers
|
|
|
|
/// described by \p CopyInfos. Append resulting instructions to \p Copies.
|
|
|
|
void TailDuplicator::appendCopies(MachineBasicBlock *MBB,
|
|
|
|
SmallVectorImpl<std::pair<unsigned,RegSubRegPair>> &CopyInfos,
|
|
|
|
SmallVectorImpl<MachineInstr*> &Copies) {
|
|
|
|
MachineBasicBlock::iterator Loc = MBB->getFirstTerminator();
|
|
|
|
const MCInstrDesc &CopyD = TII->get(TargetOpcode::COPY);
|
|
|
|
for (auto &CI : CopyInfos) {
|
|
|
|
auto C = BuildMI(*MBB, Loc, DebugLoc(), CopyD, CI.first)
|
|
|
|
.addReg(CI.second.Reg, 0, CI.second.SubReg);
|
|
|
|
Copies.push_back(C);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2016-04-08 20:35:01 +00:00
|
|
|
/// Remove the specified dead machine basic block from the function, updating
|
|
|
|
/// the CFG.
|
Codegen: Tail-duplicate during placement.
The tail duplication pass uses an assumed layout when making duplication
decisions. This is fine, but passes up duplication opportunities that
may arise when blocks are outlined. Because we want the updated CFG to
affect subsequent placement decisions, this change must occur during
placement.
In order to achieve this goal, TailDuplicationPass is split into a
utility class, TailDuplicator, and the pass itself. The pass delegates
nearly everything to the TailDuplicator object, except for looping over
the blocks in a function. This allows the same code to be used for tail
duplication in both places.
This change, in concert with outlining optional branches, allows
triangle shaped code to perform much better, esepecially when the
taken/untaken branches are correlated, as it creates a second spine when
the tests are small enough.
Issue from previous rollback fixed, and a new test was added for that
case as well. Issue was worklist/scheduling/taildup issue in layout.
Issue from 2nd rollback fixed, with 2 additional tests. Issue was
tail merging/loop info/tail-duplication causing issue with loops that share
a header block.
Issue with early tail-duplication of blocks that branch to a fallthrough
predecessor fixed with test case: tail-dup-branch-to-fallthrough.ll
Differential revision: https://reviews.llvm.org/D18226
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@283934 91177308-0d34-0410-b5e6-96231b3b80d8
2016-10-11 20:36:43 +00:00
|
|
|
void TailDuplicator::removeDeadBlock(
|
|
|
|
MachineBasicBlock *MBB,
|
2017-06-07 23:53:32 +00:00
|
|
|
function_ref<void(MachineBasicBlock *)> *RemovalCallback) {
|
2016-04-08 20:35:01 +00:00
|
|
|
assert(MBB->pred_empty() && "MBB must be dead!");
|
|
|
|
DEBUG(dbgs() << "\nRemoving MBB: " << *MBB);
|
|
|
|
|
Codegen: Tail-duplicate during placement.
The tail duplication pass uses an assumed layout when making duplication
decisions. This is fine, but passes up duplication opportunities that
may arise when blocks are outlined. Because we want the updated CFG to
affect subsequent placement decisions, this change must occur during
placement.
In order to achieve this goal, TailDuplicationPass is split into a
utility class, TailDuplicator, and the pass itself. The pass delegates
nearly everything to the TailDuplicator object, except for looping over
the blocks in a function. This allows the same code to be used for tail
duplication in both places.
This change, in concert with outlining optional branches, allows
triangle shaped code to perform much better, esepecially when the
taken/untaken branches are correlated, as it creates a second spine when
the tests are small enough.
Issue from previous rollback fixed, and a new test was added for that
case as well. Issue was worklist/scheduling/taildup issue in layout.
Issue from 2nd rollback fixed, with 2 additional tests. Issue was
tail merging/loop info/tail-duplication causing issue with loops that share
a header block.
Issue with early tail-duplication of blocks that branch to a fallthrough
predecessor fixed with test case: tail-dup-branch-to-fallthrough.ll
Differential revision: https://reviews.llvm.org/D18226
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@283934 91177308-0d34-0410-b5e6-96231b3b80d8
2016-10-11 20:36:43 +00:00
|
|
|
if (RemovalCallback)
|
|
|
|
(*RemovalCallback)(MBB);
|
|
|
|
|
2016-04-08 20:35:01 +00:00
|
|
|
// Remove all successors.
|
|
|
|
while (!MBB->succ_empty())
|
|
|
|
MBB->removeSuccessor(MBB->succ_end() - 1);
|
|
|
|
|
|
|
|
// Remove the block.
|
|
|
|
MBB->eraseFromParent();
|
|
|
|
}
|